Thrust 2: Machine Learning Models
Lead: Chao Yang
Co-Leads: Nick Bauman and Sutanay Choudhury
Goal: Formulate artificial neural network techniques that can effectively improve the performance of electronic structure methods across spatial and temporal scales.
Industrial applications of computational chemistry usually involve electronic structure methods or molecular dynamics (MD) simulations, considered workhorse techniques. Work from TEC4 aims to enhance the accuracy and broaden the applicability of these tools to enable a wider range of chemical inquiry. The team will develop deep neural network models and combine them with novel theoretical formulations. Workflows to train the models will be available to users in a high-performance computing framework. When fully trained, the models will be part of complex workflows designed to provide advanced computational chemistry services and support.
2.1 Development of Machine Learning (ML) Tools for Computational Chemistry Models
ML techniques have already been employed in computational chemistry applications, showing promise for increasing the speed and decreasing the cost of calculations. The team will explore multiple ML techniques to enable the modeling required across thrust 2. These techniques include deep neural networks, cloud-based ML tools, and reinforcement learning. The ML models will be trained via heterogeneous computing workflows with a range of hardware configurations and latency requirements. New algorithms will be developed to enable adequate scheduling and resource allocation in these workflows, enabling robust training and development. The ML models will also be trained on the cloud, with a multi-level workflow that works with extreme scales. This will prepare a robust set of models for a wide range of applications.
2.2 Development of a Compressed Form of the Effective Interactions Using Neural Networks
Recent advances in coupled cluster theory have indicated that exact energies of correlated systems could be obtained using effective Hamiltonians, simplified quantum mechanical representations of a system’s energy. TEC4 work will train deep neural networks to evaluate the interactions that define these Hamiltonians and approximate their energies. These calculations are typically very computationally expensive, and the incorporation of ML models may help provide an alternative approach to solving complex multi-body problems.
2.3 Development of High-Accuracy Interatomic Neural Network Potentials
Interatomic neural network potentials can greatly accelerate electronic structure simulations, broadening the systems that can be simulated. However, this technique is still in its early stages of study with many open questions that need to be addressed before it becomes a common practice. A major outstanding issue is how computationally expensive the simulations are, limiting the available data for training the neural network potentials. Researchers will develop workflows to more effectively train neural network potentials in a cloud environment, generating best practices for this emerging field.