The Materials Discovery Pipeline: From Hypothesis to Computation
How we structure the computational pipeline for materials discovery — from hypothesis generation through simulation, validation, and experimental hand-off.
Materials discovery is, in principle, a search problem over an astronomically large space. The number of chemically plausible inorganic compounds has been estimated at 10¹⁰⁰ or higher; the number that have been synthesised and characterised is in the millions. The ratio defines the opportunity.
In practice, discovery is constrained by the cost of evaluation. DFT calculations for a single candidate structure take minutes to hours. Synthesis and characterisation take weeks. Any discovery pipeline that does not aggressively reduce the number of candidates evaluated will be dominated by the cost of those evaluations.
Our materials discovery pipeline is organised in three stages, each designed to reduce the candidate set before passing it to the next.
Stage 1: Hypothesis generation
The first stage is generative. Starting from a target property — ionic conductivity above a given threshold, a specific bandgap, mechanical stiffness within a range — we use a generative model trained on known inorganic structures to sample candidate compositions and crystal symmetries that are, in the model's view, plausible.
The generative model does not know chemistry. It knows statistical patterns over structure space as represented in available databases. This is a useful starting point, not a reliable oracle. Its output is a ranked list of candidates, each with a predicted property estimate and an uncertainty.
Candidates below a confidence threshold are discarded. Those that survive proceed to Stage 2.
Stage 2: Quantum-informed screening
The second stage applies more expensive but more accurate evaluation. For electronic structure properties — formation energy, bandgap, redox potential — we use a combination of DFT and, for the highest-value candidates, VQE-based quantum simulation to compute corrections to the DFT baseline.
The role of quantum simulation here is targeted. We do not run VQE on all candidates; that would be prohibitively expensive and largely redundant. Instead, we use it for candidates where DFT is known to fail — strongly correlated systems, materials with d- and f-block elements, transition states in catalytic cycles. For these, VQE-derived energies anchor a correction model applied more broadly.
This stage reduces the candidate list by roughly an order of magnitude. Survivors proceed to Stage 3.
Stage 3: Synthesis recommendation
The third stage is not computational in the traditional sense. It takes the filtered candidate list and asks: which of these can plausibly be synthesised, and by what route?
This is currently the weakest link in most computational materials pipelines. Theoretical thermodynamic stability does not guarantee synthetic accessibility. We use a combination of retrosynthetic reasoning (adapted from organic chemistry tools) and experimental analogy from the literature to rank candidates by synthetic plausibility.
Output from Stage 3 is a short list — typically five to twenty candidates — delivered to experimental collaborators with predicted properties, confidence intervals, and suggested synthesis routes.
Current focus
Our current materials discovery work is concentrated on solid-state electrolytes for lithium-ion batteries and on single-atom catalysts for nitrogen fixation. Both problems share the characteristic that small changes in composition produce large changes in the target property — exactly the regime where careful quantum-informed screening pays off.
We expect to publish initial results from this program later this year.