My path to becoming a DOE scientist [Slides]

Benjamin Nebgen2022 · DOI

In a recent presentation titled "My path to becoming a DOE scientist," Benjamin Nebgen, a staff scientist at Los Alamos National Laboratory (LANL), outlined a transformative approach to computational materials science. By moving beyond the traditional constraints of quantum chemistry, Nebgen’s work demonstrates how machine learning (ML) can bridge the gap between high-accuracy quantum simulations and large-scale classical dynamics.

Presentation title slide for "My path to becoming a DOE scientist" by Benjamin Nebgen.

The Accuracy vs. Scale Trade-off

Computational chemistry and physics have long been defined by a fundamental trade-off between accuracy and computational scale. Traditional Quantum Mechanics (QM) methods, such as solving the Schrödinger equation, provide exceptional accuracy but suffer from poor scaling—typically $O(N^3)$, where $N$ represents the number of electrons. This computational bottleneck restricts QM simulations to very small systems of a few hundred atoms.

Conversely, classical force fields ($F=ma$) are computationally efficient and can simulate millions of atoms. However, they often lack the chemical accuracy required to model complex phase transitions or the behavior of novel materials under extreme conditions.

Machine Learning as a Third Pillar

Nebgen proposes Machine Learning (ML) interatomic potentials as a "third pillar" that combines the best of both worlds. These models "learn" the energy surface of a system from high-fidelity quantum data, allowing for large-scale simulations that maintain quantum-level accuracy at a fraction of the cost.

Active Learning Workflow

A key innovation in this approach is the use of Active Learning (AL) to automate the generation of training data. Rather than manually selecting configurations, an automated system explores the "phase space" of a material. It utilizes an ensemble of ML potentials to identify regions where the model’s uncertainty is high.

When "bad data coverage" is detected, the system automatically triggers a high-fidelity Quantum Espresso (QM) calculation to generate new ground-truth data. This uncertainty-driven sampling minimizes human bias and labor, allowing the model to "ask" for data only when it is necessary.

High-Performance Computing and Software Integration

The development of these potentials requires significant computational power and specialized software. Nebgen’s team utilizes several open-source tools developed at LANL:

  • HIPPYNN: A library for building machine-learned interatomic potentials.
  • PYSEQM: A PyTorch-based tool for solving reduced-order Hamiltonian models for electron motion.
  • pyDRESCALk: A code for tensor factorization used in AI platforms.

The scale of this work is exemplified by the use of the Summit supercomputer at Oak Ridge National Laboratory. Leveraging Summit’s 27,648 V100 GPUs, the team generated a massive dataset for Aluminum to train potentials capable of matching experimental properties, such as melting temperatures across various pressure ranges (up to 250 GPa).

Key Results: Modeling Aluminum

The effectiveness of this ML-driven approach is demonstrated in the modeling of Aluminum's melting curve. The ML potential developed via active learning successfully identified various phases—including Liquid, Face-Centered Cubic (FCC), and Body-Centered Cubic (BCC)—under extreme pressure, closely matching experimental data.

This work has received significant recognition, including a 2021 R&D 100 award for the team's contributions to Artificial Intelligence and Tensor Factorization platforms. Furthermore, the technical success of these methods has secured a $1.6M/year project focused on advancing ML interatomic potentials for multi-element systems.

Future Directions

While the results for Aluminum are promising, the transition to more complex, multi-element systems remains a primary challenge. Future work will focus on reducing the "human-in-the-loop" requirement for initial parameter setup and addressing the high computational cost of the initial training phase. As high-performance computing (HPC) resources continue to evolve, these ML potentials are poised to become a standard tool for exploring the frontiers of materials science.

Summary updated

Abstract

Abstract not provided.
BibTeX
@misc{nebgen2022my,
  author = {Benjamin Nebgen},
  title = {{My path to becoming a DOE scientist [Slides]}},
  year = {2022},
  doi = {10.2172/1845243},
}
Figure