1st Student-Focused SciML Symposium @ GT 2023
Virtual Symposium
16th, 21st & 30th Nov, 2023
16th, 21st & 30th Nov, 2023
| Home | Registration | Program |
| 8:00-8:05 AM ET | Introductions - The Symposium begins |
| All participants and speakers are welcomed to the first Sci-ML symposium! |
| 8:05-8:20 AM ET | Enhancing Option Pricing with Neural ODEs and SDEs |
| Team Sharpe Explorers |
| 8:20-8:35 AM ET | Enhancing Epidemic Forecasting Through Hybrid ML and Epidemiological Models |
| Team Py-Iguana |
| 8:35-8:50 AM ET | Scientific Machine Learning for Generalized Drug-Target Interaction Prediction |
| Team Polaris |
| 8:50-9:15 AM ET | Staying grounded: scientific machine learning with physical inductive biases |
| Jack Richter-Powell (Massachusetts Institute of Technology) | |
| Applying machine learning methods to problems in the physical sciences has proven successful in a variety of different settings. Specifically, there have been a number of cases where ML methods have yielded more accurate results than conventional hand developed numerical methods - a far from exhaustive set of examples includes: protein folding (AlphaFold), weather forecasting (FourCastNet among others) and simulation of quantum states in computational chemistry (Ferminet/DeepQMC). These advancements have demonstrated the efficacy of applying machine learning in these domains, but questions remains over the validity of these learned approaches when compared to their classical counterparts. Namely, many classical numerical solvers can be shown to enforce certain governing physical principles by construction, and often it is also possible to identify their failure modes by analyzing the mathematical structure of their algorithms. By contrast, ML and more specifically deep learning methods, come without these guarantees, as well as proving harder if not impossible to analyze. In this talk, we'll examine a few of my past and ongoing works bridging this gap by incorporating principled inductive biases into machine learning models. We'll discuss how this combination allows us to enjoy the increased power and flexibility of modern machine learning without sacrificing all the benefits classical numerical methods provide. |
| 8:10-8:35 AM ET | Federated scientific machine learning for approximating functions and solving differential equations |
| Handi Zhang (University of Pennsylvania) | |
| Neural networks have revolutionized the domain of scientific machine learning (SciML), offering innovative approaches to tackle complex problems governed by partial differential equations (PDEs). In real-world applications, training datasets that these data-drive methods rely on may not store in the centralized format, and the concerns of data privacy may result in communi- cation inefficiency and leave the potential improvement of collaboratively trained deep learning models. Federated Learning (FL), as a decentralized approach that collaboratively trains the global model with reserved data privacy, addresses the issues of isolated data pools and sensitive data concerns. This talk delves into the integration of federated learning with scientific machine learning to approximate complex functions and solve differential equations. First, we proposed various data assignment methods to distribute datasets across local clients to control non-iid levels in synthetic data under regression setting. Subsequently, we quantified the level of data heterogeneity in approximating functions and learning PDE solutions by utilizing the 1-Wasserstein distance and systematically exploring the relationship between non-iidness and performance of federated models. Experiments encompass various PDEs, including both inverse and forward problems in 1D and 2D, and demonstrates that incorporation of federated learning with PINNs and DeepONet preserves satisfactory accuracy and promotes efficient parallel training. |
| 8:35-8:45 AM ET | Increasing Efficiency of DFT with Surrogate Models of Graph Embeddings |
| Team Efficient DFT |
| 8:50-9:05 AM ET | Innovating Autonomous Robotics with Dual-Task Machine Learning for Navigation and Stability |
| Team Stardust |
| 9:05-9:15 AM ET | Solar Flare Peak Emission Flux Prediction |
| Team AutoCV |
| 8:10-8:35 AM ET | DeltaRCWA: a PEDS-driven solver for metamaterial scattering surrogates |
| Lorenzo Xavier Van Munoz (Massachusetts Institute of Technology) | |
| At the heart of modern optics, solving Maxwell's equations is a common routine in computational inverse design. Many algorithms for designing the macroscopic flow of light rely on fast solvers for the equations resolved at sub-wavelength scales to capture intricately patterned nanostructures. Since the design space of the 2D or 3D structure has so many free parameters, we were motivated to use a machine learning approach known as physics-enhanced deep surrogates (PEDS) to learn the forward scattering problem. We present DeltaRCWA, a Maxwell solver for scattering of surface impedances, and use it to learn the scattering of a multilayered metamaterial design. |
| 8:35-8:50AM ET | A Scientific Machine Learning Framework for Neutron Star Gravitational Modeling |
| Team Cassiopeia |
| 8:50-9:05 AM ET | Enhancing Weather Predictions with Hybrid Data and Physics-Based Models |
| Team CliML |
| 9:10-9:15 AM ET | Conclusion |
| With that, we end our first Sci-ML symposium! Thank you for everyone who participated. May we meet again!! |