Levent Burak Kara

Abstract Artificial intelligence (AI)-driven surrogate modeling has emerged as an effective alternative to physics-based simulations for 3D design, analysis, and manufacturing. These models use data-driven techniques to predict physical quantities that traditionally require computationally expensive simulations. However, the scarcity of labeled CAD-to-simulation datasets has motivated the development of self-supervised and foundation models, in which geometric representation learning is performed offline and later adapted to downstream tasks using limited labeled data. While promising, existing approaches often struggle in applications that require accurate preservation of fine-scale geometric details. This work introduces a self-supervised geometric representation learning method designed to capture fine-scale geometric features from non-parametric 3D models. Unlike traditional end-to-end surrogate models, the proposed approach decouples geometric feature extraction from downstream physics prediction by learning a latent representation guided solely by geometric reconstruction losses. Key components include near-zero-level signed distance field sampling and a batch-adaptive attention-weighted loss function, which together enhance sensitivity to subtle yet physically influential geometric variations. The proposed method is validated through two case studies involving high-dimensional design parameter regression, achieving coefficients of determination exceeding 0.98, as well as structural mechanics tasks that demonstrate strong few-shot prediction performance for reaction forces and deformation fields. Comparisons with parametric surrogate models further illustrate our method's ability to bridge geometric and physics-based representations, providing an effective surrogate modeling solution in data-scarce settings.

Semantic Scene Completion (SSC) from monocular RGB images is a fundamental yet challenging task due to the inherent ambiguity of inferring occluded 3D geometry from a single view. While feed-forward methods have made progress, they often struggle to generate plausible details in occluded regions and preserve the fundamental spatial relationships of objects. Such accurate generative reasoning capability for the entire 3D space is critical in real-world applications. In this paper, we present FlowSSC, the first generative framework applied directly to monocular semantic scene completion. FlowSSC treats the SSC task as a conditional generation problem and can seamlessly integrate with existing feed-forward SSC methods to significantly boost their performance. To achieve real-time inference without compromising quality, we introduce Shortcut Flow-matching that operates in a compact triplane latent space. Unlike standard diffusion models that require hundreds of steps, our method utilizes a shortcut mechanism to achieve high-fidelity generation in a single step, enabling practical deployment in autonomous systems. Extensive experiments on SemanticKITTI demonstrate that FlowSSC achieves state-of-the-art performance, significantly outperforming existing baselines.

Abstract Metamaterials are engineered materials composed of specially designed unit cells that exhibit extraordinary properties beyond those of natural materials. Complex engineering tasks often require heterogeneous unit cells to accommodate spatially varying property requirements. However, designing heterogeneous metamaterials poses significant challenges due to the enormous design space and strict compatibility requirements between neighboring cells. Traditional concurrent multiscale design methods require solving an expensive optimization problem for each unit cell and often suffer from discontinuities at cell boundaries. On the other hand, data-driven approaches that assemble structures from a fixed library of microstructures are limited by the dataset and require additional postprocessing to ensure seamless connections. In this work, we propose a neural network-based metamaterial design framework that learns a continuous two-scale representation of the structure, thereby jointly addressing these challenges. Central to our framework is a multiscale neural representation in which the neural network takes both global (macroscale) and local (microscale) coordinates as inputs, outputting an implicit field that represents multiscale structures with compatible unit cell geometries across the domain, without the need for a predefined dataset. We use a compatibility loss term during training to enforce connectivity between adjacent unit cells. Once trained, the network can produce metamaterial designs at arbitrarily high resolution, hence enabling infinite upsampling for fabrication or simulation. We demonstrate the effectiveness of the proposed approach on mechanical metamaterial design, negative Poisson’s ratio, and mechanical cloaking problems with potential applications in robotics, bioengineering, and aerospace.

Aeroacoustic emissions from intense evaporation are widely measured yet often treated as noisy byproducts and used mainly in empirical monitoring. Here, we show that airborne sound encodes physics-governed sub-millisecond fingerprints of vapor-jet dynamics in excessive vaporization, exemplified by vapor keyholes in laser metal processing. From first principles, we develop a vapor-jet-cavity oscillation framework and incorporate it into an aeroacoustic formulation, thereby coupling measured sound to transient cavity depth and oscillation frequency. Reconciled with synchronized multimodal in-situ data, airborne acoustics enable accurate tracking of vapor-cavity properties within tens to hundreds of microseconds. Combined with newly discovered correlations, cavity-jet-acoustic theory recasts the transition from steady, pore-free to pore-shedding vaporizations as a critical-frequency event. Aeroacoustic emissions thus become scalable, physics-guided, and cost-efficient probes of rapidly evolving liquid-vapor systems.

Aeroacoustic emissions from intense evaporation are widely measured yet often treated as noisy byproducts and used mainly in empirical monitoring. Here, we show that airborne sound encodes physics-governed sub-millisecond fingerprints of vapor-jet dynamics in excessive vaporization, exemplified by vapor keyholes in laser metal processing. From first principles, we develop a vapor-jet-cavity oscillation framework and incorporate it into an aeroacoustic formulation, thereby coupling measured sound to transient cavity depth and oscillation frequency. Reconciled with synchronized multimodal in-situ data, airborne acoustics enable accurate tracking of vapor-cavity properties within tens to hundreds of microseconds. Combined with newly discovered correlations, cavity-jet-acoustic theory recasts the transition from steady, pore-free to pore-shedding vaporizations as a critical-frequency event. Aeroacoustic emissions thus become scalable, physics-guided, and cost-efficient probes of rapidly evolving liquid-vapor systems.

Designing cryoprotectant agent (CPA) cocktails for vitrification is challenging because formulations must be concentrated enough to suppress ice formation yet non-toxic enough to preserve cell viability. This tradeoff creates a large, multi-objective design space in which traditional discovery is slow, often relying on expert intuition or exhaustive experimentation. We present a data-efficient framework that accelerates CPA cocktail design by combining high-throughput screening with an active-learning loop based on multi-objective Bayesian optimization. From an initial set of measured cocktails, we train probabilistic surrogate models to predict concentration and viability and quantify uncertainty across candidate formulations. We then iteratively select the next experiments by prioritizing cocktails expected to improve the Pareto front, maximizing expected Pareto improvement under uncertainty, and update the models as new assay results are collected. Wet-lab validation shows that our approach efficiently discovers cocktails that simultaneously achieve high CPA concentrations and high post-exposure viability. Relative to a naive strategy and a strong baseline, our method improves dominated hypervolume by 9.5\% and 4.5\%, respectively, while reducing the number of experiments needed to reach high-quality solutions. In complementary synthetic studies, it recovers a comparably strong set of Pareto-optimal solutions using only 30\% of the evaluations required by the prior state-of-the-art multi-objective approach, which amounts to saving approximately 10 weeks of experimental time. Because the framework assumes only a suitable assay and defined formulation space, it can be adapted to different CPA libraries, objective definitions, and cell lines to accelerate cryopreservation development.

Designing cryoprotectant agent (CPA) cocktails for vitrification is challenging because formulations must be concentrated enough to suppress ice formation yet non-toxic enough to preserve cell viability. This tradeoff creates a large, multi-objective design space in which traditional discovery is slow, often relying on expert intuition or exhaustive experimentation. We present a data-efficient framework that accelerates CPA cocktail design by combining high-throughput screening with an active-learning loop based on multi-objective Bayesian optimization. From an initial set of measured cocktails, we train probabilistic surrogate models to predict concentration and viability and quantify uncertainty across candidate formulations. We then iteratively select the next experiments by prioritizing cocktails expected to improve the Pareto front, maximizing expected Pareto improvement under uncertainty, and update the models as new assay results are collected. Wet-lab validation shows that our approach efficiently discovers cocktails that simultaneously achieve high CPA concentrations and high post-exposure viability. Relative to a naive strategy and a strong baseline, our method improves dominated hypervolume by 9.5\% and 4.5\%, respectively, while reducing the number of experiments needed to reach high-quality solutions. In complementary synthetic studies, it recovers a comparably strong set of Pareto-optimal solutions using only 30\% of the evaluations required by the prior state-of-the-art multi-objective approach, which amounts to saving approximately 10 weeks of experimental time. Because the framework assumes only a suitable assay and defined formulation space, it can be adapted to different CPA libraries, objective definitions, and cell lines to accelerate cryopreservation development.

Semantic Scene Completion (SSC) from monocular RGB images is a fundamental yet challenging task due to the inherent ambiguity of inferring occluded 3D geometry from a single view. While feed-forward methods have made progress, they often struggle to generate plausible details in occluded regions and preserve the fundamental spatial relationships of objects. Such accurate generative reasoning capability for the entire 3D space is critical in real-world applications. In this paper, we present FlowSSC, the first generative framework applied directly to monocular semantic scene completion. FlowSSC treats the SSC task as a conditional generation problem and can seamlessly integrate with existing feed-forward SSC methods to significantly boost their performance. To achieve real-time inference without compromising quality, we introduce Shortcut Flow-matching that operates in a compact triplane latent space. Unlike standard diffusion models that require hundreds of steps, our method utilizes a shortcut mechanism to achieve high-fidelity generation in a single step, enabling practical deployment in autonomous systems. Extensive experiments on SemanticKITTI demonstrate that FlowSSC achieves state-of-the-art performance, significantly outperforming existing baselines.

Semantic Scene Completion (SSC) from monocular RGB images is a fundamental yet challenging task due to the inherent ambiguity of inferring occluded 3D geometry from a single view. While feed-forward methods have made progress, they often struggle to generate plausible details in occluded regions and preserve the fundamental spatial relationships of objects. Such accurate generative reasoning capability for the entire 3D space is critical in real-world applications. In this paper, we present FlowSSC, the first generative framework applied directly to monocular semantic scene completion. FlowSSC treats the SSC task as a conditional generation problem and can seamlessly integrate with existing feed-forward SSC methods to significantly boost their performance. To achieve real-time inference without compromising quality, we introduce Shortcut Flow-matching that operates in a compact triplane latent space. Unlike standard diffusion models that require hundreds of steps, our method utilizes a shortcut mechanism to achieve high-fidelity generation in a single step, enabling practical deployment in autonomous systems. Extensive experiments on SemanticKITTI demonstrate that FlowSSC achieves state-of-the-art performance, significantly outperforming existing baselines.

Keyhole-induced (KH) porosity, which arises from unstable vapor cavity dynamics under excessive laser energy input, remains a significant challenge in laser powder bed fusion (LPBF). This study presents an integrated experimental and data-driven framework using airborne acoustic emission (AE) to achieve high-resolution quantification of KH porosity. Experiments conducted on an LPBF system involved in situ acquisition of airborne AE and ex situ porosity imaging via X-ray computed tomography (XCT), synchronized spatiotemporally through photodiode signals with submillisecond precision. We introduce KHLineNum , a spatially resolved porosity metric defined as the number of KH pores per unit scan length, which serves as a physically meaningful indicator of the severity of KH porosity in geometries and scanning strategies. Using AE scalogram data and scan speed, we trained a lightweight convolutional neural network to predict KHLineNum with millisecond-scale temporal resolution, achieving an R 2 value exceeding 0.8. Subsequent analysis identified the 35 – 45 kHz frequency band of AE as particularly informative, consistent with known KH oscillations. Beyond defect quantification, the framework also enables AE-driven direct inference of KH regime boundaries on the power–velocity process map, offering a noninvasive and scalable component to labor-intensive post-process techniques such as XCT. We believe this framework advances AE-based monitoring in LPBF, providing a pathway toward improved quantifiable defect detection and process control.

Keyhole-induced (KH) porosity, which arises from unstable vapor cavity dynamics under excessive laser energy input, remains a significant challenge in laser powder bed fusion (LPBF). This study presents an integrated experimental and data-driven framework using airborne acoustic emission (AE) to achieve high-resolution quantification of KH porosity. Experiments conducted on an LPBF system involved in situ acquisition of airborne AE and ex situ porosity imaging via X-ray computed tomography (XCT), synchronized spatiotemporally through photodiode signals with submillisecond precision. We introduce KHLineNum, a spatially resolved porosity metric defined as the number of KH pores per unit scan length, which serves as a physically meaningful indicator of the severity of KH porosity in geometries and scanning strategies. Using AE scalogram data and scan speed, we trained a lightweight convolutional neural network to predict KHLineNum with millisecond-scale temporal resolution, achieving an R-squared value exceeding 0.8. Subsequent analysis identified the 35-45 kHz frequency band of AE as particularly informative, consistent with known KH oscillations. Beyond defect quantification, the framework also enables AE-driven direct inference of KH regime boundaries on the power-velocity process map, offering a noninvasive and scalable component to labor-intensive post-process techniques such as XCT. We believe this framework advances AE-based monitoring in LPBF, providing a pathway toward improved quantifiable defect detection and process control.

The quadratic complexity of self-attention limits the scalability of transformers on long sequences. We introduce Fast Low-rank Attention Routing Engine (FLARE), a token-mixing operator that realizes low-rank attention by routing information through a small set of latent tokens. Each layer induces an input-input token mixing matrix of rank at most $M$ via a minimal encode-decode factorization implemented using only two standard scaled dot-product attention (SDPA) calls. Because the dominant ${O}(NM)$ computation is expressed purely in terms of standard SDPA, FLARE is compatible with fused attention kernels and avoids materializing $M\times N$ projection matrices. FLARE further assigns disjoint latent slices to each attention head, yielding a mixture of head-specific low-rank pathways. Empirically, FLARE scales to one-million-point unstructured meshes on a single GPU, achieves state-of-the-art accuracy on PDE surrogate benchmarks, and outperforms general-purpose efficient-attention methods on the Long Range Arena suite. We additionally release a large-scale additive manufacturing benchmark dataset. Our code is available at https://github.com/vpuri3/FLARE.py.

Abstract Metamaterials are engineered materials composed of specially designed unit cells that exhibit extraordinary properties beyond those of natural materials. Complex engineering tasks often require heterogeneous unit cells to accommodate spatially varying property requirements. However, designing heterogeneous metamaterials poses significant challenges due to the enormous design space and strict compatibility requirements between neighboring cells. Traditional concurrent multiscale design methods require solving an expensive optimization problem for each unit cell and often suffer from discontinuities at cell boundaries. On the other hand, data-driven approaches that assemble structures from a fixed library of microstructures are limited by the dataset and require additional post-processing to ensure seamless connections. In this work, we propose a neural network-based metamaterial design framework that learns a continuous two-scale representation of the structure, thereby jointly addressing these challenges. Central to our framework is a multiscale neural representation in which the neural network takes both global (macroscale) and local (microscale) coordinates as inputs, outputting an implicit field that represents multiscale structures with compatible unit cell geometries across the domain, without the need for a predefined dataset. We use a compatibility loss term during training to enforce connectivity between adjacent unit cells. Once trained, the network can produce metamaterial designs at arbitrarily high resolution, hence enabling infinite upsampling for fabrication or simulation. We demonstrate the effectiveness of the proposed approach on mechanical metamaterial design, negative Poisson’s ratio, and mechanical cloaking problems with potential applications in robotics, bioengineering, and aerospace.

Abstract AI-driven surrogate modeling has become an increasingly effective alternative to physics-based simulations for 3D design, analysis, and manufacturing. These models leverage data-driven methods to predict physical quantities traditionally requiring computationally expensive simulations. However, the scarcity of labeled CAD-to-simulation datasets has driven recent advancements in self-supervised and foundation models, where geometric representation learning is performed offline and later fine-tuned for specific downstream tasks. While these approaches have shown promise, their effectiveness is limited in applications requiring fine-scale geometric detail preservation. This work introduces a self-supervised geometric representation learning method designed to capture fine-scale geometric features from non-parametric 3D models. Unlike traditional end-to-end surrogate models, this approach decouples geometric feature extraction from downstream physics tasks, learning a latent space embedding guided by geometric reconstruction losses. Key elements include the essential use of near-zero level sampling and the innovative batch-adaptive attention-weighted loss function, which enhance the encoding of intricate design features. The proposed method is validated through case studies involving high-dimensional design parameter regression, achieving an R2 score greater than 0.98 in all cases, as well as structural mechanics tasks demonstrating strong few-shot physics prediction performance. Comparisons with traditional parametric surrogate modeling highlight its potential to bridge the gap between geometric and physics-based representations, providing an effective solution for surrogate modeling in data-scarce scenarios.

Abstract A long-standing challenge in multi-scale structural design is ensuring proper connectivity between the constituent cells as each cell is being optimized toward its theoretical performance limit. We propose a new method for multi-scale topology optimization that seamlessly ensures compatibility across neighboring microstructure cells without the need for a catalog of unit cells and associated connectivity rules. Our approach consists of a topology neural network that encodes each cell’s microstructure as well as the distribution of the optimized cells within the design domain as a continuous and differentiable density field. Each cell is optimized based on a prescribed elasticity tensor. The neural network takes as input the local coordinates within a cell to represent the density distribution within a cell, as well as the global coordinates of each cell to generate spatially varying microstructure cells. As such, our approach models an n-dimensional multi-scale optimization problem as a 2n-dimensional inverse homogenization problem using neural networks. During the inverse homogenization of each unit cell, we extend the boundary of each cell by scaling the input coordinates such that the boundaries of neighboring cells are blended and, thus, jointly optimized. This enables inverse homogenization of the cells to be consistently connected without introducing discontinuities at the cell boundaries. We demonstrate our method through the design and optimization of several graded multi-scale structures.

AI-driven surrogate modeling has become an increasingly effective alternative to physics-based simulations for 3D design, analysis, and manufacturing. These models leverage data-driven methods to predict physical quantities traditionally requiring computationally expensive simulations. However, the scarcity of labeled CAD-to-simulation datasets has driven recent advancements in self-supervised and foundation models, where geometric representation learning is performed offline and later fine-tuned for specific downstream tasks. While these approaches have shown promise, their effectiveness is limited in applications requiring fine-scale geometric detail preservation. This work introduces a self-supervised geometric representation learning method designed to capture fine-scale geometric features from non-parametric 3D models. Unlike traditional end-to-end surrogate models, this approach decouples geometric feature extraction from downstream physics tasks, learning a latent space embedding guided by geometric reconstruction losses. Key elements include the essential use of near-zero level sampling and the innovative batch-adaptive attention-weighted loss function, which enhance the encoding of intricate design features. The proposed method is validated through case studies in structural mechanics, demonstrating strong performance in capturing design features and enabling accurate few-shot physics predictions. Comparisons with traditional parametric surrogate modeling highlight its potential to bridge the gap between geometric and physics-based representations, providing an effective solution for surrogate modeling in data-scarce scenarios.

A method that allows a fast and accurate registration of digital tissue models obtained during preoperative, diagnostic imaging with those captured intraoperatively using lower-fidelity ultrasound imaging techniques is presented. Minimally invasive surgeries are often planned using preoperative, high-fidelity medical imaging techniques such as MRI and CT imaging. While these techniques allow clinicians to obtain detailed 3D models of the surgical region of interest (ROI), various factors such as physical changes to the tissue, changes in the body's configuration, or apparatus used during the surgery may cause large, non-linear deformations of the ROI. Such deformations of the tissue can result in a severe mismatch between the preoperatively obtained 3D model and the real-time image data acquired during surgery, potentially compromising surgical success. To overcome this challenge, this work presents a new approach for predicting intraoperative soft tissue deformations. The approach works by simply tracking the displacements of a handful of fiducial markers or analogous biological features embedded in the tissue, and produces a 3D deformed version of the high-fidelity ROI model that registers accurately with the intraoperative data. In an offline setting, we use the finite element method to generate deformation fields given various boundary conditions that mimic the realistic environment of soft tissues during a surgery. To reduce the dimensionality of the 3D deformation field involving thousands of degrees of freedom, we use an autoencoder neural network to encode each computed deformation field into a short latent space representation, such that a neural network can accurately map the fiducial marker displacements to the latent space. Our computational tests on a head and neck tumor, a kidney, and an aorta model show prediction errors as small as 0.5 mm. Considering that the typical resolution of interventional ultrasound is around 1 mm and each prediction takes less than 0.5 s, the proposed approach has the potential to be clinically relevant for an accurate tracking of soft tissue deformations during image-guided surgeries.

Advances in CAD (Computer Aided Design) and CAM (Computer Aided Engineering) have enabled engineers and design teams to digitally design parts with unprecedented ease. Software solutions now come with a range of modules for optimizing designs for performance requirements, generating instructions for manufacturing, and digitally tracking the entire process from design to procurement in the form of product life-cycle management tools. However, existing solutions force design teams and corporations to take a primarily serial approach where manufacturing and procurement decisions are largely contingent on design, rather than being an integral part of the design process. In this work, we propose a new approach to part making where design, manufacturing, and supply chain requirements and resources can be jointly considered and optimized. We present the Generative Manufacturing compiler that accepts as input the following: (1) An engineering part requirements specification that includes quantities such as loads, domain envelope, mass, and compliance, (2) A business part requirements specification that includes production volume, cost, and lead time, (3) Contextual knowledge about the current manufacturing state such as availability of relevant manufacturing equipment, materials, and workforce, both locally and through the supply chain. Based on these factors, the compiler generates and evaluates manufacturing process alternatives and the optimal derivative designs that are implied by each process, and enables a user guided iterative exploration of the design space. As part of our initial implementation of this compiler, we demonstrate the effectiveness of our approach on examples of a cantilever beam problem and a rocket engine mount problem and showcase its utility in creating and selecting optimal solutions according to the requirements and resources. • A new approach to part making is presented. • Design, manufacturing and supply chain requirements are jointly considered to generate optimum parts. • Transparent trade-offs are revealed. • The approach ensures seamless adaptation of part design to disruptions in supply chain.

Reduced order modeling lowers the computational cost of solving PDEs by learning a low-dimensional spatial representation from data and dynamically evolving these representations using manifold projections of the governing equations. The commonly used linear subspace reduced-order models (ROMs) are often suboptimal for problems with a slow decay of Kolmogorov n -width, such as advection-dominated fluid flows at high Reynolds numbers. There has been a growing interest in nonlinear ROMs that use state-of-the-art representation learning techniques to accurately capture such phenomena with fewer degrees of freedom. We propose smooth neural field ROM (SNF-ROM), a nonlinear reduced order modeling framework that combines grid-free reduced representations with Galerkin projection. The SNF-ROM architecture constrains the learned ROM trajectories to a smoothly varying path, which proves beneficial in the dynamics evaluation when the reduced manifold is traversed in accordance with the governing PDEs. Furthermore, we devise robust regularization schemes to ensure the learned neural fields are smooth and differentiable. This allows us to compute physics-based dynamics of the reduced system nonintrusively with automatic differentiation and evolve the reduced system with classical time-integrators. SNF-ROM leads to fast offline training as well as enhanced accuracy and stability during the online dynamics evaluation. Numerical experiments reveal that SNF-ROM is able to accelerate the full-order computation by up to 199×. We demonstrate the efficacy of SNF-ROM on a range of advection-dominated linear and nonlinear PDE problems where we consistently outperform state-of-the-art ROMs. • We present a novel reduced order model based on smooth neural fields (SNF-ROM). • SNF-ROM reduces the order of PDE problems to their intrinsic manifold dimension. • The online dynamics evaluation projects the governing PDE onto the ROM manifold. • The dynamics evaluation is robust to numerical perturbations by design. • We achieve a speed-up of up to 199× over the full-order computation.

Abstract Designing for manufacturing poses significant challenges in part due to the computation bottleneck of Computer-Aided Manufacturing (CAM) simulations. Although deep learning as an alternative offers fast inference, its performance is dependently bounded by the need for abundant training data. Representation learning, particularly through pre-training, offers promise for few-shot learning, aiding in manufacturability tasks where data can be limited. This work introduces VIRL, a Volume-Informed Representation Learning approach to pre-train a 3D geometric encoder. The pretrained model is evaluated across four manufacturability indicators obtained from CAM simulations: subtractive machining (SM) time, additive manufacturing (AM) time, residual von Mises stress, and blade collisions during Laser Power Bed Fusion process. Across all case studies, the model pre-trained by VIRL shows substantial enhancements in generalizability, as measured by R2 regression results, with improved performance on limited data and superior predictive accuracy with larger datasets. Regarding deployment strategy, case-specific phenomenon exists where finetuning VIRL-pretrained models adversely affects AM tasks with limited data but benefits SM time prediction. Moreover, the efficacy of Low-rank adaptation (LoRA), which balances between probing and finetuning, is explored. LoRA shows stable performance akin to probing with limited data, while achieving a higher upper bound than probing as data size increases, without the computational costs of finetuning. Furthermore, static normalization of manufacturing indicators consistently performs well across tasks, while dynamic normalization enhances performance when a reliable task dependent input is available.

The discovery and study of new material systems rely on molecular simulations that often come with significant computational expense. We propose MDDM, a Molecular Dynamics Diffusion Model, which is capable of predicting a valid output conformation for a given input pair potential function. After training MDDM on a large dataset of molecular dynamics self-assembly results, the proposed model can convert uniform noise into a meaningful output particle structure corresponding to an arbitrary input potential. The model's architecture has domain-specific properties built-in, such as satisfying periodic boundaries and being invariant to translation. The model significantly outperforms the baseline point-cloud diffusion model for both unconditional and conditional generation tasks.

Advanced liver preservation strategies could revolutionize liver transplantation by extending preservation time, thereby allowing for broader availability and better matching of transplants. However, developing new cryopreservation protocols requires exploration of a complex design space, further complicated by the scarcity of real human livers to experiment upon. We aim to create computational models of the liver to aid in the development of new cryopreservation protocols. Towards this goal, we present an approach for generating 3D models of the liver vasculature by building upon the space colonization algorithm. Additionally, we introduce the concept of a super lobule which enables a computational abstraction of biological liver lobules. User-tunable parameters allow for vasculatures of varying depth and topology to be generated. In each model, we solve for a common lumped resistance value assigned to the super lobules, allowing the overall physiological blood pressure and flow rate through the liver to be preserved. We demonstrate our approach's ability to maintain consistency between models of varying depth. Finally, we simulate steady state machine perfusion of the generated models and demonstrate how they can be used to quickly test the effect of different boundary conditions when designing organ preservation protocols.

Recent advancements in 3D structure-based molecular generative models have shown promise in expediting the hit discovery process in drug design. Despite their potential, efficiently generating a focused library of candidate molecules that exhibit both effective interactions and structural diversity at a large scale remains a significant challenge. Moreover, current studies often lack comprehensive comparisons to high-throughput virtual screening methods, resulting in insufficient evaluation of their effectiveness. In this study, we introduce Topology Molecular Type assignment (TopMT-GAN), a novel approach using Generative Adversarial Networks (GANs) for direct structure-based design. TopMT-GAN employs a two-step strategy: constructing 3D molecular topologies within a protein pocket with one GAN, followed by atom and bond type assignment with a second GAN. This integrated approach enables TopMT-GAN to efficiently generate diverse and potent ligands with precise 3D poses for specific protein pockets. When tested on five diverse protein pockets, TopMT-GAN exhibits promising and robust performance, demonstrating a potential enrichment of up to 46 000 fold compared to traditional high-throughput virtual screening methods. This highlights its potential as a powerful tool in early-stage drug discovery, such as hit and lead generation.

Abstract In design and manufacturing engineering, the manual interpretation of 2D technical drawings remains a significant bottleneck, impeding efficiency in tasks such as part quotation and process manufacturing. While computer vision advancements have shown promise in interpreting natural images, the translation of engineering drawings into actionable data poses unique challenges. Our research proposes a novel data-driven framework that leverages computational vision and graph learning to automate the vectorization and low-level interpretation of raster engineering drawings, bringing us closer to a CAD format. Our approach involves the vectorization of rasterized engineering drawings into a collection of lines and arcs. We construct a connectivity graph from these lines and arcs and train a graph convolutional neural network to classify entity types accurately. We concentrate on 9 distinct classes: Lines, Arrowheads, Text, Dimension Lines, Center Lines, Arcs, Hidden Lines and Contours, and Misc. We achieved a notably high performance of 98.2% validation accuracy, a 20.66% improvement compared to prior methodologies, and a maximum of 46.48% increase compared to baseline learning methods. By streamlining the interpretation of technical drawings, our framework offers a pathway to enhance efficiency in modern manufacturing processes, reducing reliance on manual labor and boosting overall productivity.

Abstract Laser powder bed fusion is a mainstream additive manufacturing technology widely used to manufacture complex parts in prominent sectors, including aerospace, biomedical, and automotive industries. However, during the printing process, the presence of an unstable vapor depression can lead to a type of defect called keyhole porosity, which is detrimental to the part quality. In this study, we developed an effective approach to locally detect the generation of keyhole pores during the printing process by leveraging machine learning and a suite of optical and acoustic sensors. Simultaneous synchrotron x-ray imaging allows the direct visualization of pore generation events inside the sample, offering high-fidelity ground truth. A neural network model adopting SqueezeNet architecture using single-sensor data was developed to evaluate the fidelity of each sensor for capturing keyhole pore generation events. Our comparative study shows that the near infrared images gave the highest prediction accuracy, followed by 100 kHz and 20 kHz microphones, and the photodiode sensitive to processing laser wavelength had the lowest accuracy. Using a single sensor, over 90% prediction accuracy can be achieved with a temporal resolution as short as 0.1 ms. A data fusion scheme was also developed with features extracted using SqueezeNet neural network architecture and classification using different machine learning algorithms. Our work demonstrates the correlation between the characteristic optical and acoustic emissions and the keyhole oscillation behavior, and thereby provides strong physics support for the machine learning approach.

Abstract Machine-learned surrogate models to accelerate lengthy computer simulations are becoming increasingly important as engineers look to streamline the product design cycle. In many cases, these approaches offer the ability to predict relevant quantities throughout a geometry, but place constraints on the form of the input data. In a world of diverse data types, a preferred approach would not restrict the input to a particular structure. In this paper, we propose topology-agnostic graph U-Net (TAG U-Net), a graph convolutional network that can be trained to input any mesh or graph structure and output a prediction of a target scalar field at each node. The model constructs coarsened versions of each input graph and performs a set of convolution and pooling operations to predict the node-wise outputs on the original graph. By training on a diverse set of shapes, the model can make strong predictions, even for shapes unlike those seen during training. A 3D additive manufacturing dataset is presented, containing laser powder bed fusion simulation results for thousands of parts. The model is demonstrated on this dataset, and it performs well, predicting both 2D and 3D scalar fields with a median R2>0.85 on test geometries.

Machine-learned surrogate models to accelerate lengthy computer simulations are becoming increasingly important as engineers look to streamline the product design cycle. In many cases, these approaches offer the ability to predict relevant quantities throughout a geometry, but place constraints on the form of the input data. In a world of diverse data types, a preferred approach would not restrict the input to a particular structure. In this paper, we propose Topology-Agnostic Graph U-Net (TAG U-Net), a graph convolutional network that can be trained to input any mesh or graph structure and output a prediction of a target scalar field at each node. The model constructs coarsened versions of each input graph and performs a set of convolution and pooling operations to predict the node-wise outputs on the original graph. By training on a diverse set of shapes, the model can make strong predictions, even for shapes unlike those seen during training. A 3-D additive manufacturing dataset is presented, containing Laser Powder Bed Fusion simulation results for thousands of parts. The model is demonstrated on this dataset, and it performs well, predicting both 2-D and 3-D scalar fields with a median R-squared > 0.85 on test geometries. Code and datasets are available online.

Advances in CAD and CAM have enabled engineers and design teams to digitally design parts with unprecedented ease. Software solutions now come with a range of modules for optimizing designs for performance requirements, generating instructions for manufacturing, and digitally tracking the entire process from design to procurement in the form of product life-cycle management tools. However, existing solutions force design teams and corporations to take a primarily serial approach where manufacturing and procurement decisions are largely contingent on design, rather than being an integral part of the design process. In this work, we propose a new approach to part making where design, manufacturing, and supply chain requirements and resources can be jointly considered and optimized. We present the Generative Manufacturing compiler that accepts as input the following: 1) An engineering part requirements specification that includes quantities such as loads, domain envelope, mass, and compliance, 2) A business part requirements specification that includes production volume, cost, and lead time, 3) Contextual knowledge about the current manufacturing state such as availability of relevant manufacturing equipment, materials, and workforce, both locally and through the supply chain. Based on these factors, the compiler generates and evaluates manufacturing process alternatives and the optimal derivative designs that are implied by each process, and enables a user guided iterative exploration of the design space. As part of our initial implementation of this compiler, we demonstrate the effectiveness of our approach on examples of a cantilever beam problem and a rocket engine mount problem and showcase its utility in creating and selecting optimal solutions according to the requirements and resources.

Abstract Scalar fields, such as stress or temperature fields, are often calculated in shape optimization and design problems in engineering. For complex problems where shapes have varying topology and cannot be parametrized, data-driven scalar field prediction can be faster than traditional finite element methods. However, current data-driven techniques to predict scalar fields are limited to a fixed grid domain, instead of arbitrary mesh structures. In this work, we propose a method to predict scalar fields on arbitrary meshes. It uses a convolutional neural network whose feature maps at multiple resolutions are interpolated to node positions before being fed into a multilayer perceptron to predict solutions to partial differential equations at mesh nodes. The model is trained on finite element von Mises stress fields, and once trained, it can estimate stress values at each node on any input mesh. Two shape datasets are investigated, and the model has strong performance on both, with a median R2 value of 0.91. We also demonstrate the model on a temperature field in a heat conduction problem, where its predictions have a median R2 value of 0.99. Our method provides a potential flexible alternative to finite element analysis in engineering design contexts. Code and datasets are available online.

Designing for manufacturing poses significant challenges in part due to the computation bottleneck of Computer-Aided Manufacturing (CAM) simulations. Although deep learning as an alternative offers fast inference, its performance is dependently bounded by the need for abundant training data. Representation learning, particularly through pre-training, offers promise for few-shot learning, aiding in manufacturability tasks where data can be limited. This work introduces VIRL, a Volume-Informed Representation Learning approach to pre-train a 3D geometric encoder. The pretrained model is evaluated across four manufacturability indicators obtained from CAM simulations: subtractive machining (SM) time, additive manufacturing (AM) time, residual von Mises stress, and blade collisions during Laser Power Bed Fusion process. Across all case studies, the model pre-trained by VIRL shows substantial enhancements on demonstrating improved generalizability with limited data and superior performance with larger datasets. Regarding deployment strategy, case-specific phenomenon exists where finetuning VIRL-pretrained models adversely affects AM tasks with limited data but benefits SM time prediction. Moreover, the efficacy of Low-rank adaptation (LoRA), which balances between probing and finetuning, is explored. LoRA shows stable performance akin to probing with limited data, while achieving a higher upper bound than probing as data size increases, without the computational costs of finetuning. Furthermore, static normalization of manufacturing indicators consistently performs well across tasks, while dynamic normalization enhances performance when a reliable task dependent input is available.

Reduced order modeling lowers the computational cost of solving PDEs by learning a low-order spatial representation from data and dynamically evolving these representations using manifold projections of the governing equations. While commonly used, linear subspace reduced-order models (ROMs) are often suboptimal for problems with a slow decay of Kolmogorov $n$-width, such as advection-dominated fluid flows at high Reynolds numbers. There has been a growing interest in nonlinear ROMs that use state-of-the-art representation learning techniques to accurately capture such phenomena with fewer degrees of freedom. We propose smooth neural field ROM (SNF-ROM), a nonlinear reduced modeling framework that combines grid-free reduced representations with Galerkin projection. The SNF-ROM architecture constrains the learned ROM trajectories to a smoothly varying path, which proves beneficial in the dynamics evaluation when the reduced manifold is traversed in accordance with the governing PDEs. Furthermore, we devise robust regularization schemes to ensure the learned neural fields are smooth and differentiable. This allows us to compute physics-based dynamics of the reduced system nonintrusively with automatic differentiation and evolve the reduced system with classical time-integrators. SNF-ROM leads to fast offline training as well as enhanced accuracy and stability during the online dynamics evaluation. Numerical experiments reveal that SNF-ROM is able to accelerate the full-order computation by up to $199\times$. We demonstrate the efficacy of SNF-ROM on a range of advection-dominated linear and nonlinear PDE problems where we consistently outperform state-of-the-art ROMs.

A long-standing challenge is designing multi-scale structures with good connectivity between cells while optimizing each cell to reach close to the theoretical performance limit. We propose a new method for direct multi-scale topology optimization using neural networks. Our approach focuses on inverse homogenization that seamlessly maintains compatibility across neighboring microstructure cells. Our approach consists of a topology neural network that optimizes the microstructure shape and distribution across the design domain as a continuous field. Each microstructure cell is optimized based on a specified elasticity tensor that also accommodates in-plane rotations. The neural network takes as input the local coordinates within a cell to represent the density distribution within a cell, as well as the global coordinates of each cell to design spatially varying microstructure cells. As such, our approach models an n-dimensional multi-scale optimization problem as a 2n-dimensional inverse homogenization problem using neural networks. During the inverse homogenization of each unit cell, we extend the boundary of each cell by scaling the input coordinates such that the boundaries of neighboring cells are combined. Inverse homogenization on the combined cell improves connectivity. We demonstrate our method through the design and optimization of graded multi-scale structures.

With a growing demand for high-quality fabrication, the interest in real-time process and defect monitoring of laser powder bed fusion (LPBF) has increased, leading manufacturers to incorporate a variety of online sensing methods including acoustic sensing, photodiode sensing, and high-speed imaging. However, real-time acquisition of high-resolution melt pool images in particular remains computationally demanding in practice due to the high variability of melt pool morphologies and the limitation of data caching and transfer, making it challenging to detect the local lack-of-fusion (LOF) defect occurrences. In this work, we propose a new acoustic and thermal information-based monitoring method that can robustly infer critical LPBF melt pool morphological features in image forms and detect spatially-dependent LOF defects within a short time period. We utilize wavelet scalogram matrices of acoustic and photodiode clip data to identify and predict highly time-resolved (within a 1.0 ms window) visual melt pool characteristics via a well-trained data-driven pipeline. With merely the acoustic and photodiode-collected thermal emission data as the input, the proposed pipeline enables data-driven inference and tracking of highly variable melt pool visual characteristics with R2≥0.8. We subsequently validate our proposed approach to infer local LOF defects between two adjacent scanlines, showing that our proposed approach can outperform our selected baseline theoretical model based on previous literature. Revealing the physical correlation between airborne acoustic emission, thermal emission, and melt pool morphology, our work demonstrates the feasibility of creating an efficient and cost-effective acoustic- and thermal-based approach to facilitate online visual melt pool characterization and LOF defect detection. We believe that our work can further contribute to the advances in quality control for LPBF.

Abstract We introduce a computational approach for the design of target-specific peptides. Our method integrates a Gated Recurrent Unit-based Variational Autoencoder with Rosetta FlexPepDock for peptide sequence generation and binding affinity assessment. Subsequently, molecular dynamics simulations are employed to narrow down the selection of peptides for experimental assays. We apply this computational strategy to design peptide inhibitors that specifically target β -catenin and NF- κ B essential modulator. Among the twelve β -catenin inhibitors, six exhibit improved binding affinity compared to the parent peptide. Notably, the best C-terminal peptide binds β -catenin with an IC 50 of 0.010 ± 0.06 μM, which is 15-fold better than the parent peptide. For NF- κ B essential modulator, two of the four tested peptides display substantially enhanced binding compared to the parent peptide. Collectively, this study underscores the successful integration of deep learning and structure-based modeling and simulation for target specific peptide design.