LM3: Learning Multiscale Material Models

Department of Mechanical Engineering, University of Michigan, Ann Arbor
Thomas D Swinburne, PhD
tswin-at-umich-dot-edu

Metallic alloys uniquely combine strength, ductility and reusability, properties which emerge from an often mysterious hierarchy of plasticity mechanisms. Alloy design requires inverting complex structure-processing-property relationships from sparse and indirect data, a significant challenge for human or AI scientists.

The LM3 lab develops new simulation methods connecting atomic dynamics, microstructure and mechanical properties. We collaborate widely and use ideas from applied math, machine learning, physics and mechanics. A recent focus is harnessing high-dimensional latent spaces for multiscale inverse design and uncertainty quantification.

We are hiring in Ann Arbor & Paris!
See open positions or email tswin-at-umich-dot-edu. Lab members have gone on to obtain competitive positions in academia and industry.

Current research topics
Projects balance theory/applications depending on your interest.

• Application focus:
  • Slip transfer in complex dislocation networks
  • High-dpa irradiation damage for nuclear fusion
  • Prediction of phase diagrams across the periodic table
  • Mechanistic inverse design of non-dilute alloys
• Theory/method development focus:
  • Model-form uncertainty quantification (UQ) for ML surrogates
  • Data-driven forecasting of long-timescale behavior
  • End-to-end differentiable simulations for UQ and design
  • Inverse fine-tuning of atomic machine learning models

Selected recent work
See Google Scholar for an up-to-date list.

• Score matched free energies (Nat. Comm. 2025 & AI4Mat / code)
  MLIP phase diagrams as functions, for UQ and inverse problems !
• Entropy of dislocation glide (Nat. Comm. 25/PAFI/Rodney Group)
  Why ML potentials are required to accurately model bcc plastic flow.
• Implicit differentiation in MD (NPJ 2025 featured / code)
  MLIP energy landscapes as functions, for UQ and inverse problems !
• Misspecification uncertainty in MD (NPJ 2025 / code)
  Using POPS to explore model-form UQ of SNAP and MACE.
  
• POPS deterministic model-form UQ (ML:S&T 2025 / code)
  Solving Bayes’ ignorance of model-form UQ (misspecification).
• MACE foundation potentials (Csyani Group / JCP 2025)
  Section A.13. Dislocations are a tough extrapolation test for UMLIPs.
• Coarse-graining & forecasting atomic simulations ( PRL 2023)
  ML feature vectors used for MLIPs can also forecast complex futures
• Embedding ab initio in MLIPs for solute studies (Acta Mat. 2023 )
  QM/ML for dislocation-solute interactions. Applied to W alloys
  
• A15 defects in fcc (Nat. Comm. 2023 / TAMMBER / Marinica Group)
  Irradiation defects can grow as three-dimensional Laves phase clusters

Recent or upcoming travel and talks
EL2026, Telluride, USA, 07/26
SIAM UQ, Minneapolis, USA, 03/26
Frontiers of Theory and AI in Sustainable Mat. Sci., Dusseldorf, 01/26
USACM UQ, Webinar, USA, 01/26
NeurIPS AI4Mat, San Diego, USA, 12/25
Dislocations 2025, Miami, USA, 11/25
CECAM: UQ from DFT to ML, EPFL, Lausanne, 11/25
CoMPASs workshop, ICMS, Edinburgh, 11/25 (unavailable)
IPAM Electrochemistry, UCLA, 10/25
USACM Nanomechanics 2025, Urbana-Champaign, USA, 09/25