Scope |
Numerical simulations have enabled a new paradigm in materials discovery. However, they face various challenges, including: (i) high computational cost, which usually prevents the simulation of large systems over extended timescales, (ii) limited accuracy (e.g., due to lack of reliable interatomic forcefields), and (iii) difficulties when it comes to the inverse design optimization of materials (simulations are often not differentiable). On the other hand, artificial intelligence and machine learning offer a promising pathway for materials modeling and accelerated discovery of new materials with exceptional properties. However, machine learning models also face some limitations as they: (i) rely on the existence of large, consistent, and accurate datasets, (ii) excel at interpolating materials’ properties but tend to have challenges with extrapolations, (iii) by solely relying on data, can violate the laws of physics and chemistry, and (iv) typically offer limited interpretability. In that regard, data-driven machine learning models and knowledge-driven high-fidelity simulations have the potential to inform, advance, and complement each other—and to address each other’s deficiencies. This symposium builds on the idea that the lack of intimate integration between data- and knowledge-driven modeling is a missed opportunity in materials science.
This symposium will explore new modeling approaches that seamlessly combine and integrate machine learning and simulations—wherein simulation informs machine learning, machine learning advance simulations, or closed-loop integrations thereof. It will bring together experts in numerical simulations and machine learning, both from Academia, National Laboratories, and Industry.
Topics of interest include, but are not limited to:
- Multi-fidelity models, data-fusion, and transfer learning approaches
- Machine learning to inform simulations (e.g., machine-learned interatomic forcefields)
- Physics-informed machine learning and symbolic learning
- "Self-driving" simulations, reinforcement learning, and active learning
- Graph neural networks for materials modeling
- Automatic differentiation, inverse problems, and deep generative models
- Machine learning for “finding needles in haystacks” in simulation output data
- Rare events sampling and automated identification of collective variables
- Machine learning for structural and topology optimization
- Development of machine-learned surrogate simulators
- Natural language processing for materials modeling
- Use of hardware dedicated to deep learning (e.g., TPUs) to accelerate simulations |