Atoms to AI: Building a Quantum Capacitor Surrogate with Python and GNNs

How can an independent researcher explore the future of energy storage without access to an expensive physical lab? We turn to Python. This session explores the development of a Graph Neural Network (GNN) surrogate model designed to predict the Quantum Capacitance of 2D material heterostructures—specifically a "sandwich" of Graphene and hexagonal Boron Nitride (h-BN).

By utilizing open-source libraries and public datasets, we walk through a complete end-to-end pipeline to simulate "virtual experiments":

• Data Ingestion: Using pymatgen to pull high-fidelity atomic structures from the Materials Project database.
• Geometric Assembly: Leveraging ASE (Atomic Simulation Environment) to programmatically construct 2-2-2 layered stacks and modulate interlayer distances at the angstrom scale.
• The Model: Building a PyTorch-based GNN that learns the relationship between 3D atomic coordinates and predicted capacitance.
• The Result: Developing a surrogate model that provides immediate feedback on design variations—mapping 3D coordinates to target values—without the need for physical trial and error.

Attendees will leave with a practical look at how Python empowers researchers to build their own "virtual labs," bridging the gap between material science and machine learning for rapid iteration in solid-state device design.

Presented by

Terry Howald