fromatomstobuildings,themathbreaks

Problem statement

No AI or machine learning method can bridge atomic-to-continuum scales with provable accuracy guarantees. Predicting macroscopic material properties from atomic-scale physics requires coupling simulations across ~10 orders of magnitude in length and time — from quantum mechanics (angstroms, femtoseconds) through molecular dynamics (nanometers, nanoseconds) to continuum mechanics (meters, seconds). AI surrogate models trained on fine-grained simulation data can approximate coarse-grained dynamics, but no method provides provable error bounds on this coarse-graining, and no framework guarantees that critical rare events (phase transitions, crack nucleation, defect migration) are preserved in the surrogate.

Why this matters

Multi-scale modeling is essential for materials design, drug discovery, climate modeling, and engineering simulation. The Materials Genome Initiative has invested $500+ million since 2011 to accelerate materials discovery, but multi-scale prediction remains the central bottleneck — designing a new alloy or polymer still requires decades of iterative experiment because simulation cannot reliably predict bulk properties from atomic composition. A reliable AI multi-scale framework would compress materials development timelines from decades to years, with implications across aerospace, energy, medicine, and manufacturing.

What’s been tried and why it hasn’t worked

Machine learning interatomic potentials (MLIPs — GAP, NequIP, MACE) accurately reproduce ab initio energies and forces but only at the atomistic scale — they don't bridge to continuum. Coarse-grained molecular dynamics with ML force fields can run larger systems faster but loses information about rare events and fails when the coarse-graining scheme encounters conditions outside its training distribution. Physics-informed neural networks (PINNs) solve PDEs but don't learn the PDEs — they require the governing equations to be known, which is exactly what multi-scale modeling tries to discover. Graph neural networks for materials property prediction (CGCNN, MEGNet) correlate structure to properties but are purely data-driven with no physics-based error bounds. The fundamental challenge is that coarse-graining is an irreversible information-losing projection — reconstructing fine-grained behavior from coarse-grained representations is an ill-posed inverse problem.

What would unlock progress

Mathematical theory for provably accurate coarse-graining — determining what information is necessarily lost in scale transitions and bounding the resulting prediction error. Data-efficient methods for learning rare-event dynamics from molecular simulations without requiring prohibitively long trajectories. Hybrid frameworks that couple physics-based models at each scale with learned scale-bridging operators, validated against experiment at each level.

Entry points for student teams

A student team could take a well-studied system (e.g., water, a simple metal) where both atomistic and continuum behavior are well-characterized, train an ML surrogate model on molecular dynamics data, and systematically measure where the surrogate fails — identifying the types of phenomena that are lost in coarse-graining. Alternatively, a team could compare different ML interatomic potentials' predictions of a specific material property (e.g., thermal conductivity) against experimental measurements, quantifying the accuracy gap between atomic-scale ML and macroscopic observation. Relevant skills: computational materials science, machine learning, molecular dynamics, statistical mechanics.