Why the Most Advanced Materials Teams Don't Start in the Lab

In most engineering disciplines, simulation precedes physical construction.

Structural engineers model before they build. Aeronautical engineers run computational fluid dynamics before a prototype enters a wind tunnel. The principle is consistent: reduce the cost of being wrong by doing as much as possible in computation before committing physical resources.

Materials development has not followed this principle - not because it was the preferred approach, but because the tools were not sufficient. Computational methods were either too slow or too expensive to run at meaningful scale. Physical experimentation was the only practical option.

That constraint is no longer absolute.

It is now possible to screen thousands of candidate materials computationally - evaluating composition, structure, stability and performance properties - before committing to physical synthesis. Development cycles that have historically consumed years can be compressed to weeks. Physical testing resources concentrate on high-confidence candidates rather than a wide field of unknowns.

This is not a marginal improvement. It is a change in where the process begins. Organisations operating simulation-first are not simply working faster. They are making better decisions earlier, with less capital at risk.

That is the structural advantage computational materials infrastructure creates.

The simulation-first methodology rests on a layered computational stack. At the foundation sits ab initio (first principles) calculation, specifically DFT using electronic structure software such as FHI-aims, which provides electronic structure data from which higher-level models are trained. This data is the reference against which all subsequent approximations are validated.

Above this, machine learning interatomic potentials are trained to reproduce the DFT potential energy surface at significantly reduced computational cost. The critical metric is parity validation: the degree to which MLIP predictions reproduce DFT results across a range of configurations, including those outside the training distribution. Without rigorous parity validation, the simulation layer cannot be trusted for screening decisions that will subsequently drive physical investment.

The practical workflow proceeds as follows. A candidate search space is defined by composition range, structural constraints and target properties. High-throughput generation tools - including special quasirandom structures (SQS) for alloy modelling and structure search algorithms - populate this space with candidate configurations. SQS were introduced by Zunger, Wei, Ferreira and Bernard to efficiently represent the correlation functions of a random alloy in small periodic supercells.¹ Results are filtered, ranked and presented as a reduced set of high-confidence candidates for physical validation.

The computational layer does not replace physical validation. It narrows the search space before physical validation begins. The final arbiters of material performance remain physical tests under real operating conditions. Simulation changes the quality of candidates that arrive at that stage.