AI-Assisted Device Design
When designing devices, we try to generate designs to optimize some objective function, such as computation time, power draw, spectrum, etc. In the past, solving design problems relied on the designers intuition, searching large design spaces or even sampling NP-Hard problems. To reduce the design problem's difficulty, we use latent models, such as LDMs, GANs, and VAEs, to extract symmetries and features in the data to construct a more efficient search space for design optimizers.
Physics-Inspired Generative Modeling
Physics-inspired generative modeling leverages principles from statistical mechanics, electromagnetism, and other physical theories to build efficient and interpretable probabilistic models. By explicitly incorporating physical symmetries, constraints, and known laws into generative models, we significantly reduce the complexity and dimensionality of the learning problem. This approach enables accurate data generation, improved generalization, and deeper insights into underlying physical processes, making it especially powerful for applications in simulation, material science, and system prediction.
Optimization
Optimization techniques aim to efficiently search large parameter spaces and solve complex, often NP-hard problems. We leverage advanced algorithms including Markov Chain Monte Carlo (MCMC) and Markov processes to explore solution spaces systematically, improving convergence and performance across applications in computational physics, machine learning, and device design.