Solid-state electrolyte development has a scale problem that doesn't get discussed honestly enough. The compositional space for sulfide, oxide, and halide electrolytes is enormous. Just within the Li-P-S-X (X = halogen) system, the number of synthesizable compositions at meaningful quality is in the thousands when you account for dopant substitutions and processing variations. An experimental team synthesizing and characterizing 8-12 samples per week would need decades to cover that space systematically.
The traditional response is chemical intuition plus combinatorial synthesis: a senior researcher identifies a promising sub-region of composition space, the team makes a grid of samples across that region, and results from one grid inform the next. It works, but it's slow and it's heavily biased by existing knowledge. If the optimal composition is in a region that looks unpromising by conventional intuition, you won't find it.
Prediction-driven screening doesn't replace experimental intuition. It changes the economics of when you commit synthesis resources, and it can surface candidates outside the intuition-informed search region.
What the composition space looks like for solid electrolytes
Take the argyrodite family as a concrete case. The parent structure Li6PS5X (X = Cl, Br, I) has been studied extensively. But substitution possibilities are broad: halide site mixing (Cl/Br mixtures), cation substitutions at the Li sites, partial replacement of P with Ge or Si, and oxygen incorporation at the S sites. Each substitution dimension adds a continuous variable. A systematic 5-point grid over 6 substitution dimensions gives 5^6 = 15,625 candidate compositions. That's a decade of synthesis at typical lab throughput.
The prediction problem is: given composition, predict room-temperature ionic conductivity (often reported as log(sigma) in units of log(S/cm)) and electrochemical stability window. These are the two primary screening filters before anyone commits to pellet pressing, sintering, and EIS measurement.
The challenge is that conductivity in argyrodites is sensitive to halide site disorder, which depends on synthesis conditions, and to grain boundary density, which depends on sintering. A prediction from composition alone is inherently a prediction of the intrinsic bulk property under idealized synthesis. That's a real limitation we're upfront about with anyone using our platform for electrolyte screening.
The prediction-then-synthesis workflow
The workflow we've built for electrolyte screening has three stages, and the key is that experimental synthesis doesn't start until stage two is complete.
Stage one is virtual screening. We generate candidate compositions by systematic substitution and evaluate each candidate through a composition-to-property model trained on existing conductivity measurements. This model runs in milliseconds per candidate, so 15,000 candidates complete in under a minute. The output is a ranked list with predicted conductivity and a prediction interval for each candidate.
Stage two is filtering and prioritization. We apply a set of hard cutoffs: predicted conductivity above a threshold (say, 0.5 mS/cm), predicted stability window that doesn't preclude use with the intended anode and cathode, and a synthesis feasibility check that flags compositions with elements known to be unstable under typical sulfide processing conditions. After this filter, a candidate list of thousands shrinks to tens or low hundreds.
The prioritization within the filtered list uses an acquisition function that combines predicted performance with prediction uncertainty. A candidate that predicts well but has high uncertainty (because it's in an undersampled region of composition space) gets a boost relative to a candidate that predicts well and has low uncertainty. This is the exploration-exploitation tradeoff from Bayesian optimization, applied to the synthesis queue.
Stage three is experimental characterization. The filtered and prioritized list becomes the synthesis plan. Compared to a combinatorial grid, this synthesis campaign is much smaller and covers a non-rectangular region of composition space: the candidates are clustered where the model predicts good performance plus high information value, not on a geometric grid.
Campaign economics: a comparison
To make the economics concrete, consider a screening campaign over a 3-variable substitution space in the Li-PS-X system. A traditional full-grid approach at 5 points per dimension requires 125 synthesis-and-characterization runs. At roughly 3 working days per run (synthesis, pressing, sintering, EIS), that's approximately 375 working days to complete the grid, at a consumable cost of perhaps $80-150 per run depending on materials used.
A prediction-driven approach on the same space typically requires 20-40 experimental runs to cover the highest-value regions plus enough measurement points to update the prediction model. Total time: 60-120 working days. The savings are real, but they come with a caveat: if the model's pre-screening is poorly calibrated, you might miss a high-performing region that the model incorrectly scored low. This is why calibrated uncertainty is as important as point predictions. A model that knows what it doesn't know will flag the underexplored regions for sampling; a model with overconfident intervals will not.
Where prediction-driven screening works less well
We've seen prediction-driven screening produce weak results in two situations, and it's worth being direct about them.
First: when the training data is heavily concentrated in one sub-region of composition space. If every existing conductivity measurement comes from argyrodite structures and you're trying to screen NASICON-type candidates, the model's composition-conductivity mapping extrapolates badly. The model learned an argyrodite-specific relationship, not a general one. The uncertainty estimates will be wide (which is correct), but wide uncertainty means the acquisition function treats almost all NASICON candidates as equally worth sampling, which is just random selection with extra steps.
Second: when the target property depends critically on a variable you can't encode in the composition. Grain boundary conductance in sulfide electrolytes is strongly affected by pressing pressure and sintering temperature. Two identically composed samples processed differently can show ionic conductivities differing by an order of magnitude. A composition-only model conflates these processing differences as noise on the composition-conductivity relationship, degrading its predictions for any composition that's sensitive to processing.
We're not claiming prediction-driven screening eliminates these failure modes. We're saying it's faster and more informative than a blind grid when the model is trained on reasonably well-distributed data covering the chemistry of interest.
How measured results feed back into the model
One advantage that compounds over time: each experimental measurement from the screening campaign is a new training point for the prediction model. After completing a prediction-driven campaign, the model knows more about the composition-conductivity relationship in the specific sub-region you investigated. The next campaign in an adjacent region starts with a better-informed prior.
This is iterative learning, and it's one of the things that makes a platform approach more valuable than a one-shot prediction. The first campaign pays for the model update; the second campaign benefits from it. Over several campaigns in a given materials family, the model's accuracy in that family improves substantially, and the amount of experimental synthesis needed per candidate decreases.
For the argyrodite work we've been involved in, the accuracy improvement across three sequential campaigns was measurable: the fraction of synthesized candidates meeting the target conductivity threshold improved from roughly 25% in the first campaign to above 55% in the third, as the model accumulated better coverage of the substitution space. The model wasn't getting smarter in any fundamental sense; it just had better training data each time.
Electrolyte development timelines are long. Getting the prediction-screening workflow right early in a program compresses the campaign from years to months. The key is starting with a model that's honest about what it doesn't know, and committing to a feedback loop that makes the model more knowledgeable with each experimental batch.