Experimental design in materials and computational chemistry

How it is used

Materials chemistry asks which combination of elements, and which recipe for making them, will give a substance the properties wanted: a metal that springs back to shape, a solar absorber, a glass. The space of possible compositions and processing conditions is vast, and each attempt costs furnace time, instrument time or computer time. Experimental design here means using a model of what is already known to choose the next thing to try. A model is fitted to existing measurements or calculations, asked to predict outcomes it has not seen, and its predictions, together with how unsure it is, are used to rank the candidates. The chosen experiment is run, the result is added, and the cycle repeats.

In the examples, this takes several forms. Models pick synthesis targets and guess the ingredients and the oven temperature for each one, so that many recipes can be attempted together. In a closed loop, each round of measurement is interpreted by a model that divides a sample into regions of different crystal structure, and that interpretation guides where to measure next. In simulation, a cheap model of the forces between atoms is trained on expensive quantum-mechanical calculations, used to run simulations, and asked to flag the structures it is least equipped to handle, which are then calculated properly and added to its training set.

Why researchers use it

The reason is cost. Making and measuring a new material is slow, and the accurate calculations are slow too, so the number of attempts is small compared with the number of possibilities. Researchers reach for these methods because a fitted model, even a rough one, lets them spend those few attempts where they are most likely to teach something. Models that report their own uncertainty are favoured, since a candidate can be worth trying either because it looks good or because nobody knows what it would do.

In the record

materials-chemistryRobotic lab with machine learning proposes and runs 353 inorganic synthesis experimentsRanked known materials by similarity to each target to propose precursors, filled in missing elements, and predicted a synthesis temperature from precursor and target properties.materials-chemistryA learning algorithm ran a synchrotron beamline to find a phase-change materialSplit each round's diffraction patterns into regions of distinct crystal phases, using a graph-based method encoding physical constraints, to guide the next measurement.materials-chemistryMachine-learned force model simulates silicon-oxygen structures from glass surfaces to monoxide grainsTrained cheap models of interatomic forces on quantum calculations, ran simulations with them, and sent structures beyond their reliable range for proper calculation.materials-chemistryMachine learning picked which shape-memory alloys to make next, round by roundFitted regression models from composition-derived elemental descriptors to measured thermal hysteresis in an alloy family, with uncertainties estimated by resampling the training data.

Limits

These models only know the region they were trained on, and most materials data sits in a narrow band of compositions and recipes that already worked. A model will often look confident just outside it and be wrong, so a loop can settle into a familiar corner and never reach the interesting one. Descriptors built from elemental properties ignore microstructure, impurities and the details of handling, which frequently decide whether a synthesis succeeds. Simplifications made for convenience, such as sharing one predicted temperature across a batch, trade accuracy for throughput.

Written by AIxSci: the background from general knowledge, the examples from checked records.