About the webinar:
AI foundation models are transforming how we simulate materials. Trained on millions of quantum-mechanical calculations, universal machine-learning potentials promise near-DFT accuracy across the periodic table at a fraction of the cost. But how far can we trust them for real engineering materials?
In this webinar, Daniel Marchand (SINTEF Industry) shares lessons from benchmarking state-of-the-art foundation models on aluminium alloys, testing them on the properties that matter in practice. He will discuss where these models already deliver, where they still fail, and why training data quality often matters more than model size. You will leave with practical guidance on choosing the right model for your application, and a clearer view of what AI can realistically offer metallurgy and materials design today.
Who is the webinar for?
-
Researchers and PhD students in computational materials science, physics and chemistry who use or plan to use machine-learning interatomic potentials (MLIPs) for molecular dynamics and atomistic simulations.
-
Industrial R&D engineers and materials scientists in the metals, automotive, aerospace, energy and manufacturing sectors who want to understand whether AI-based atomistic simulation is mature enough to support alloy design and process development.
-
Developers of ML models and materials datasets, and HPC/AI practitioners, interested in benchmarking, data curation and the compute trade-offs of large models.
Basic familiarity with atomistic simulation (DFT or molecular dynamics) is helpful, but no prior experience with foundation models is required.
Key takeaways for participants:
-
A clear picture of the current landscape of universal interatomic potentials and how they differ in architecture and training data.
-
Evidence of where today’s foundation models succeed and fail for metals, from bulk and defect properties to dislocations and precipitate shearing.
-
Why DFT settings and defect coverage in training datasets strongly affect accuracy for metallic systems.
-
What fine-tuning can and cannot achieve, including the value of small, targeted datasets.
-
Practical guidance for choosing between foundation and custom potentials, weighing accuracy, GPU cost, system size and development time, and why validation against DFT remains essential.
Background reading: D. Marchand, “Foundation models for metallurgy?”, MRS Bulletin 50, 805–818 (2025). Open access: https://doi.org/10.1557/s43577-025-00911-0
Speaker bio:
Daniel Marchand is a research scientist at SINTEF Industry in Oslo, Norway, where he works on machine-learning interatomic potentials and atomistic modelling for metallurgical applications. He received his PhD from École Polytechnique Fédérale de Lausanne (EPFL) in 2022 with a thesis on neural network potentials for age-hardening aluminium alloys, developing Behler–Parrinello potentials for the Al–Cu–Mg–Zn system to study solute clustering, precipitation and dislocation–precipitate interactions. His recent work focuses on rigorous, property-driven benchmarking of universal foundation models for metals, with the goal of making AI-driven atomistic simulation reliable enough for industrial alloy development.