Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

VIRL: Volume-Informed Representation Learning towards Few-shot Manufacturability Estimation

Yu-hsuan Chen, Jonathan Cagan, Levent Burak kara

TL;DR

This work tackles the CAM simulation bottleneck in manufacturability prediction by introducing VIRL, a volume-informed self-supervised pretraining strategy for a CAD boundary-representation encoder. It pretrains the encoder via a volumetric SDF-focused task and evaluates on four CAM-derived indicators across a dataset of over 20,000 parts, demonstrating improved few-shot generalization and strong performance with larger labeled sets. The study also investigates deployment strategies, showing LoRA provides a stable, data-efficient alternative to full finetuning, and that dynamic normalization with task-dependent input boosts AM tasks when reliable heuristics exist while static normalization remains robust otherwise. Overall, VIRL offers a scalable initialization for downstream manufacturability regression, reducing data needs and speeding up early-stage design workflows.

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 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.

VIRL: Volume-Informed Representation Learning towards Few-shot Manufacturability Estimation

TL;DR

This work tackles the CAM simulation bottleneck in manufacturability prediction by introducing VIRL, a volume-informed self-supervised pretraining strategy for a CAD boundary-representation encoder. It pretrains the encoder via a volumetric SDF-focused task and evaluates on four CAM-derived indicators across a dataset of over 20,000 parts, demonstrating improved few-shot generalization and strong performance with larger labeled sets. The study also investigates deployment strategies, showing LoRA provides a stable, data-efficient alternative to full finetuning, and that dynamic normalization with task-dependent input boosts AM tasks when reliable heuristics exist while static normalization remains robust otherwise. Overall, VIRL offers a scalable initialization for downstream manufacturability regression, reducing data needs and speeding up early-stage design workflows.

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 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.
Paper Structure (19 sections, 2 equations, 12 figures, 7 tables)

This paper contains 19 sections, 2 equations, 12 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. Proposed architecture
  5. Case Studies Formulation
  6. Testing Pretrained Model on Downstream Tasks
  7. Results and Discussion
  8. Models' Performance on Various Tasks and Dataset Size
  9. Visualizing Latent Space via Dimension Reduction
  10. Exploring Probing and Finetuning Alternatives for VIRL-pretrained Model
  11. Effect of Including Task Dependent Input
  12. Ablation Study
  13. Effect of Modifying Hidden Layer Width and Latent Code Width
  14. Effect of Early Stopping
  15. Limitations and Future Work
  16. ...and 4 more sections

Figures (12)

  • Figure 1: Schematic diagram of the VIRL’s architecture. The numbers in parentheses represent the dimension of latent vectors and linear layers.
  • Figure 2: 48 variations of a part being transposed and rotated and their corresponding 5-variable augmentation code.
  • Figure 3: The loss curves during pretraining. (a) total loss (Decoder1 + Decoder2) (b) reconstruction loss (Decoder2), and (c) bounding box loss (Decoder1).
  • Figure 4: SDF reconstruction results in the test set.
  • Figure 5: Downstream regression pipeline. After the encoder is pre-trained with VIRL, the latent representation is used to regress on four manufacturability cases studies curated from simulation.
  • ...and 7 more figures