Big Ideas Take Time: From Hessian-Trained AI Potentials and Open Discovery Tools to Quantum Materials and Heavy-Element Chemistry
Chemical Engineering and Materials Science
Physics and Astronomy (affiliated)
Michigan State University
Abstract
Big ideas take time. An open tool, a new method or a more exact calculations can take years, but this is how science is done. In this seminar I will follow four such long-term bets from our group, connecting machine learning/AI, open science, materials design & relativistic quantum chemistry.
First, we teach machine-learning interatomic potentials (MLIPs) the curvature of chemistry. Conventional MLIPs learn from energies and forces, yet the Hessian, the matrix of second derivatives, encodes the local shape of the potential energy surface. Training on Hessians improves predictions for non-equilibrium structures, transition states, reaction paths and vibrational spectra, keeps molecular dynamics stable, and needs fewer training structures [1]. Because full Hessians are costly to compute and store, we introduce Projected Hessian Learning (PHL), which supervises curvature through Hessian–vector products along random directions and matches full-Hessian accuracy while training more than 24 times faster per epoch [2, 3].
Second, methods matter only when others can use them. CheMLFlow is an open-source platform that turns cheminformatics and materials-informatics studies into reproducible, auditable workflows, from data curation to scaffold-aware validation and automated reports and gives AI coding agents a structured way to run experiments under human supervision [4]. SALSA couples a data-driven substitution model, which proposes more than 30,000 new compositions, with redox-stability screening, evolutionary structure prediction and hybrid density functional theory to find earth-abundant semiconductors [5, 6]. Two open, executable e-books, the Quantum Computing Live Book [7] and the new edition of the Machine Learning and AI [8], bring these tools into the classroom.
Hybrid density functional theory (DFT) explained how electrochemical cycling restructures defective PtSe2 into a fuel-cell catalyst that, after 42,000 cycles, has 2.6 times the mass activity of commercial Pt/C [9]. We showed that terahertz pulses focused at the tip of a scanning tunneling microscope drive a picometer-scale shear of the top atomic layer of WTe2 that switches its surface topology [10]. Our first-principles screens of twisted bilayers and two-dimensional metal halides map how stacking angle and composition tune band gaps [11, 12].
Finally, we go where learned models still lack reference data: the heavy elements. Motivated by the separation of nuclear waste, we compared four-component relativistic and scalar (ZORA) calculations for an isostructural family of plutonium, americium, berkelium and californium complexes with a redox-active ligand. Four-component Dirac–Hartree–Fock agrees best with experiment, scalar ZORA-DFT is an efficient first approximation, and relativity reshapes how 5f and 6d orbitals take part in bonding [13, 14]. The same tools predicted protactinium defects in diamond as quantum sensors for physics beyond the Standard Model, which are now being tested experimentally [15].
Together, these stories make the case that patient investment in better data, open tools and exact science turn first-principles simulations into a partner for discovery (it just takes time!).
References
Bio
Dr. José L. Mendoza-Cortés is an Assistant Professor in the Department of Chemical Engineering and Materials Science at Michigan State University, with an affiliation in the Department of Physics and Astronomy. A theoretical and computational condensed-matter physicist, chemist and materials scientist, he develops quantum-mechanical, relativistic and machine-learning methods. He received his Ph.D. (2012) from Caltech. He co-authored with Prof. Yaghi a widely used review of the building units (SBU) of metal–organic frameworks (MOFs), the materials recognized by the 2025 Nobel Prize in Chemistry. His honors include two Royal Society of Chemistry Emerging Investigator selections (2018, 2025), Advanced Materials Rising Stars (2025), four consecutive Scialog Fellowships (2020–2023) and a Telly Award for a documentary on his lab. He has published 69 peer-reviewed articles, cited more than 10,000 times. He has been funded consecutively by NSF for 11 years as well as by DOE, ONR, NVIDIA, Sloan, RCSA. Through his doctoral advisor, his academic family tree reaches back to Marie Curie.
Persons with disabilities have the right to request and receive reasonable accommodation. Please call the Department of Chemical Engineering and Materials Science at (517) 355-5135 at least three days prior to the seminar; requests received after this date will be met when possible.