Materials and Process covers how a material is actually made, and the growing body of work using machine-learned potentials, generative models and closed-loop robotic search to make processing route, not just composition, the designed variable.
Summary
A material’s properties are set as much by its processing history as by its chemical formula. The same stoichiometry can be an amorphous glass or a dense crystal, a rough film or a single-crystalline one, depending on temperature ramp, flux ratio, quench rate or applied field waveform. “Materials & Process” is the part of materials science that treats those synthesis and processing variables as the object of study: growth windows, nucleation pathways, grain-boundary segregation, pore architecture, network topology in polymers.
The current research front is the attempt to put this domain under computational and automated control. Three layers are visible in recent literature. First, cheap surrogates for quantum mechanics: machine-learned interatomic potentials such as MACE and NequIP now simulate molten lithium carbonate structure and transport at near-ab-initio accuracy, and machine-learned polaron dynamics extends sampling of reduced rutile TiO2(110) from the picoseconds accessible to ab initio molecular dynamics to several nanoseconds. Second, inverse and search tools: diffusion models that recover crystal structure from powder X-ray diffraction patterns, and hybrid solid-state nudged elastic band frameworks that reach the same minimum energy pathways as DFT with up to a 7-fold speedup. Third, closed-loop experiment: evolutionary search over scanning-probe tip-bias waveforms with uncertainty-aware deep kernel learning, validated experimentally on ferroelectric thin films, and open-source software to connect AI planners to heterogeneous lab robots.
Underneath this sits conventional process craft that still delivers most of the demonstrated results: molecular beam epitaxy of Mn2In2Se5 via Mn intercalation into In2Se3, with growth windows of 250-450 degC and Se:Mn flux ratios of 1.1-3.1, yielding 1.5 nm RMS roughness and 0.35 degree rocking-curve FWHM for the MnSe overlayer; parts-per-million cobalt chain-transfer catalyst to lower crosslink density, glass transition and modulus in methacrylate nematic elastomers; carbonisation and pore-closure strategies for hard carbon sodium-ion anodes.
The parameters that decide whether the automated layer displaces the craft layer are: whether surrogate models stay accurate away from their training distribution, whether in-situ diagnostics carry enough signal to close a loop, and whether process knowledge transfers between tools and labs.
Evidence base
- 15 June 2026: a hybrid SSNEB framework combining pretrained EquiformerV2 and eSEN with DFT reached up to 7-fold speedup while converging to the same minimum energy pathways as first principles for CsPbI3, GaN and TiO2.
- 16 June 2026: all eight leading universal machine-learning interatomic potentials tested produced amorphous IrO2 densities of 1-4 g/cm3 under conventional melt-quench, against an ab initio reference of 10.04 g/cm3.
- 15 June 2026: machine-learned polaron dynamics reached several nanoseconds of sampling at oxygen-deficient rutile TiO2(110), beyond the picosecond limit of ab initio molecular dynamics, showing small-polaron mobility suppressed by several orders of magnitude versus bulk.
- 15 June 2026: MBE via Mn intercalation produced single-crystalline (0001) Mn2In2Se5 with 1.5 nm RMS roughness over growth windows of 250-450 degC and Se:Mn flux ratios of 1.1-3.1, with extended deposition giving alpha-MnSe(111) up to 8 nm and 0.35 degree rocking-curve FWHM.
- 16 June 2026: the NIMO open-source platform reported a unified interface with twelve AI algorithms driving six self-driving-lab implementations, from electrolyte discovery to fuel-cell process informatics.
- 15 June 2026: positionally controlled mechanosynthetic donation of carbon and abstraction of silicon was demonstrated with inverted-mode STM and functionalised molecular tools on a clean model build site.
Open questions
- Do the fixes for universal-potential failure under melt-quench, pressure-targeted fine-tuning and revised NVT-quench/NPT-equilibration, generalise beyond IrO2, or does each new amorphous system need its own diagnosis?
- Can in-situ diagnostics in epitaxy be extended to predict the metrics RHEED currently misses, notably surface roughness, which correlated only weakly with RHEED embeddings for GaSe?
- Does the closed-loop waveform-discovery result on ferroelectric thin films transfer to processing variables that are slower, costlier and less reversible than a scanning-probe tip bias?
- Does XRDiff’s structure recovery hold on experimental powder patterns with real peak broadening, impurity phases and background, rather than the simulated spectrum-structure pairs it is benchmarked on?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.