AiPEX Lab

Engineering simulation results stress plots, thermal gradients, streamlines, fluid flow patterns, contain domain-specific information that requires human expertise to understand and draw meaningful insights, creating a bottleneck in automated frameworks. Can AI models that excel at describing medical images and satellite photos learn to read finite element analysis? We built OpenSeeSimE, an automated benchmark with over 200,000 question-answer pairs spanning 10,000 simulations across structural, thermal, and fluid dynamics. State-of-the-art vision-language models achieve only 29-47% accuracy on engineering-specific questions. This performance gap reveals fundamental challenges in domain transfer; engineering simulations require understanding physical principles, recognizing patterns in scientific visualizations, and reasoning about multi-scale phenomena. Future work will use this benchmark to evaluate different techniques for improving AI-automated workflows, such as agent-based design processes that decompose interpretation tasks, incorporate physics-based reasoning, and keep humans in the loop while automating repetitive analysis.

Finite element analysis for complex geometries under uncertainty can take hundreds of hours of computation. Engineering designers, who design for uncertain loading conditions, need a way to see how variations in geometry, material properties, and loading conditions affect performance. These design spaces can have up to thousands of variations, but running full simulations for each is prohibitively expensive. Neural network surrogates offer a solution: train once on a representative dataset, then predict outcomes in milliseconds. We build a system that uses three-dimensional unstructured graph meshes as inputs to test if the current state-of-the-art neural network surrogate models can understand both geometry and physics. Results show engineering-grade accuracy for stochastic finite element problems is still far from reality. Future work will use this benchmark to evaluate different techniques for improving AI automated workflows via neural network surrogate models to improve efficiency and enable them to be used reliably in real-world engineering workflows.

Most reinforcement learning treats systems as black boxes—observe states, take actions, get rewards, repeat. This works for environments with clearly definable states, such as computer games, but fails for engineering design, which often spans multiple domains with unclear state and action spaces. For instance, a robotic manipulator isn't just mechanical linkages; it integrates motors (electrical), hydraulic actuators, sensors, and control systems. Defining custom state and action spaces for each design scenario quickly becomes cumbersome. Bond graphs provide a unified modeling language for these multi-domain systems, representing energy flow through mechanical, electrical, hydraulic, and thermal components using the same mathematical framework. By incorporating bond graph structure directly into deep Q-learning, we have developed Bond-DQN: an approach that leverages the universal graph representation for multi-domain engineering design. We demonstrate Bond-DQN on a realistic suspension vibration suppression task, achieving 20.5% better performance than traditional parametric optimization.

A well-fitted prosthetic limb shouldn't be a luxury. Traditional prosthetics require industrial manufacturing equipment costing hundreds of thousands of dollars, and expert prosthetists for custom fitting. We have developed a workflow that enables resource-constrained communities to design, analyze, and fabricate customized prosthetic devices using accessible 3D printing technology and smartphones. The system uses video captured on any smartphone to reconstruct accurate 3D geometry of residual limbs, applies parametric design algorithms to generate custom-fitted components, predicts mechanical performance using machine learning models trained on printing parameters, and produces fully 3D-printable designs requiring zero industrial components. This approach enables rapid iteration and designs adapted to specific activities, representing a new paradigm for democratized product design in resource-constrained settings. 3D-printed parts are highly sensitive to printing parameters. The same digital design printed at 50% infill density versus 100% infill, or at different layer heights, nozzle temperatures, and print speeds, produces parts with dramatically different mechanical properties—stiffness, strength, toughness. Running finite element analyses for every parameter combination is prohibitively time-consuming when designers need to explore hundreds of variations to optimize for conflicting objectives like strength versus weight. We have built comprehensive datasets mapping printing parameters to mechanical properties across over 10,000 test specimens and developed machine learning models that predict material behavior directly from print settings with 4-5% error. This capability transforms the design process: instead of printing multiple physical prototypes to find optimal parameters, designers can now explore parameter spaces digitally in minutes, capturing complex interactions between parameters and enabling the discovery of combinations that achieve specific performance targets.

Manufacturing systems are typically controlled by monitoring every sensor at every timestep and adjusting parameters continuously. While this approach is intuitively sound it is expensive and introduces noise from over-control which can actually harm throughput. Moreover, when reinforcement learning (RL) is applied as controllers in production line systems, such dense control paradigms become even more suboptimal not only because of the computation overhead, but also the unclear correlation between the reward signals and actions taken. Manufacturing systems exhibit natural rhythms and time constants, as products require time to be processed and transferred through the system. Effective RL-based control therefore requires intervening at the right moments rather than continuously controlling the system. We developed a method that identifies the suitable control interval using the structure and the parameters of the production line which provide RL algorithms information regarding when to observe system state and let dynamics evolve in order to apply control actions. Our approach analyzes production structure together with processing parameters to determine appropriate control intervals, enabling RL agents to acquire richer and more informative reward signals. Experimental results demonstrate improved throughput compared to baseline every-timestep control strategies, with learned policies uncovering effective performances near the heuristic optimal of the production line performance.

Standard deep reinforcement learning treats everything as patterns to be learned from scratch: which states lead to good outcomes, which actions work in which situations, how different parts of the system interact. But we already know physics—energy is conserved, momentum transfers according to Newton's laws, contact forces depend on material properties and geometry. We're developing approaches that incorporate physics-based priors directly into learning architectures: energy-based networks that ensure learned dynamics respect conservation laws, Lagrangian and Hamiltonian neural networks that parameterize physics correctly, bond graph representations for multi-domain systems, and symmetry-preserving architectures. The results are dramatic: robots learn stable gaits and manipulation strategies with 10-100× fewer samples than black-box approaches, policies generalize to novel terrains and objects not seen during training, and learned behaviors remain safe because physical constraints are built-in rather than discovered.

Robots operating in the real world don't inhabit the smooth, continuous state spaces that most learning algorithms assume. A quadruped walking doesn't just move continuously through joint angles and velocities—it experiences discrete events when feet make and break contact with the ground, suddenly changing system dynamics. These hybrid systems, with continuous motion punctuated by discrete mode switches, are notoriously difficult for standard machine learning. We're developing hybrid system identification methods that explicitly detect and model these discrete events alongside continuous dynamics. Our approach simultaneously identifies continuous dynamics within each mode (flight phase, single-support, double-support), the discrete events that trigger mode transitions (foot contact, liftoff), and the conditions under which transitions occur. This hybrid structure enables model-based reinforcement learning that's far more sample-efficient than model-free methods, achieving more stable locomotion across varied terrains while transferring across different gaits because it understands the physical structure underlying legged locomotion.