The phrase "high-throughput computational screening" appears frequently enough in materials R&D discussions that it has started to lose meaning. Teams describe workflows that range from evaluating a few dozen DFT calculations per week to genuinely large passes over tens of thousands of structures evaluated with machine-learned potentials. The number that gets cited (5,000 structures, 50,000 structures) rarely comes with the context that determines whether it means anything for the practical discovery problem at hand. This post unpacks the decisions that actually determine whether a high-throughput screening run produces useful candidates or an expensive list of noise.
Candidate Space Definition Is the Critical First Decision
How you construct the initial candidate set matters far more than the number of structures you screen. There are several approaches in common use: structure prototype enumeration, compositional interpolation within a known phase family, generative model output, and random structure search using evolutionary algorithms or random sampling in composition space.
Prototype enumeration starts from known crystal structure types (rock salt, spinel, layered oxide, perovskite variants) and systematically substitutes elements across allowed sites. For a battery cathode problem where you already know the structural class you want, this is usually the most productive starting point because it keeps you close to synthesizable chemistry. The candidate space is bounded and chemically interpretable.
Generative approaches produce more diverse structures but at the cost of including many structures that are either already well-characterized (large duplication with public databases) or far from any known synthesizable phase. In our experience with oxide ceramics and cathode materials, roughly 40 to 60 percent of generative candidate pools turn out to duplicate known MP or OQMD entries when structure fingerprinting is applied. Deduplication is a necessary preprocessing step and changes the effective size of the screening run significantly. We note this not to dismiss generative approaches, which genuinely do surface novel topologies, but because teams running these workflows sometimes count their pre-deduplicated input size as the "number of structures screened," which overstates what actually happened.
Filter Ordering Affects Throughput and Hit Rate Together
A screening pipeline is a sequence of filters, each with a different cost and a different discriminating power. The order you apply them determines how much compute you spend to reach a given number of high-quality candidates at the output. The general principle is cheap filters first, expensive filters later. But the practical version of this rule requires thinking about the false-negative rate of each cheap filter: if your early cheap filter eliminates a large fraction of actually viable candidates, you are not saving compute, you are discarding the candidates that would have been worth synthesizing.
For inorganic stability screening, a common three-stage filter sequence looks like this: first, composition-level filters that eliminate structures with clearly unfavorable elemental combinations (charge imbalance, obvious instability from known electronegativity considerations); second, structure-level MLIP energy evaluation with a convex hull cutoff, keeping candidates within, say, 80 meV/atom of the hull; third, property-target filters applied only to the survivors of the stability filter. The second filter is where MLIP speed matters most, because DFT at that scale would be prohibitive. The third filter runs on a small enough set that even moderately expensive property calculations are affordable.
The Convex Hull Cutoff Choice
The convex hull energy cutoff is a parameter that most screening workflows treat as a fixed value but that should be calibrated against the specific synthesis conditions. A 0 meV/atom cutoff means you are only passing thermodynamically stable structures at 0 K. A 100 meV/atom cutoff passes many metastable structures. For solid-state synthesis routes at high temperature, metastable phases frequently form and can be kinetically stable under working conditions. For thin-film deposition, the accessible phase space is different again.
For cathode oxide materials, structures 50 to 100 meV/atom above the convex hull are routinely synthesized and used in battery cells. A workflow with a strict 20 meV/atom cutoff will systematically exclude the novel metastable phases that might be most interesting. We are not saying loose cutoffs are always better, but the choice should be deliberate and justified relative to the synthesis method and target application, not simply set to the most conservative value because that produces a smaller output list.
What "10,000 Structures Screened" Actually Means
There are several different things a team might mean when they say they screened 10,000 structures. Walking through the distinctions clarifies what the number measures and what it does not.
- 10,000 structures from a generated or enumerated candidate pool, filtered only by composition rules, with no energy evaluation. This is a population count, not a screening result.
- 10,000 structure-energy evaluations with an MLIP, applied to a candidate pool that may have started larger. This is the genuine throughput number and the one that reflects compute investment.
- 10,000 structures from a prototype database that were retrieved and filtered by property lookup (e.g., from MP). This is a database query, not a screening run, and says nothing about your model's ability to evaluate new compositions.
- 10,000 unique compositions, where each composition may have multiple structure variants evaluated. Depending on how structures are counted, the evaluation count could be much higher.
When we describe the throughput of our screening pipeline, we mean the number of unique candidate structures for which the MLIP energy and force evaluation was actually computed within the run, after deduplication and composition pre-filtering. That is the number that scales with compute time and that should be compared to DFT compute equivalents.
How the Output List Should Be Used
A ranked candidate list from a high-throughput screening run is input to a subsequent decision, not a conclusion. The output of screening is a prioritized queue for more expensive validation, whether that is DFT, experimental characterization, or both. Teams that treat the MLIP-predicted energy ranking as a final answer, and dispatch synthesis on the top-ranked structures without further validation, are making a significant assumption about the reliability of the model for those specific compositions.
The responsible workflow is to run DFT validation on a representative sample of the MLIP shortlist before committing synthesis time. How large that sample needs to be depends on how well-characterized the training-set coverage is for the composition families in question. For well-characterized phase families (common NMC-type compositions, for example), a smaller validation fraction is defensible. For novel compositions at the edge of the training distribution, a higher fraction makes sense and the DFT validation results should feed back into the MLIP training data to improve future screening quality.
The Metric That Actually Matters
In our work, the metric we watch most closely is not total structures screened but synthesis-viable candidates per DFT-compute-hour. This ratio reflects how well the screening pipeline is using the compute budget to produce candidates that actually get made and characterized. A pipeline that screens 50,000 structures and produces 3 synthesis-viable candidates per 100 DFT-validation compute-hours is worse than a pipeline that screens 5,000 structures and produces 18 synthesis-viable candidates per 100 DFT-validation compute-hours, despite the former having a larger throughput headline number.
Optimizing for this ratio requires getting the candidate space definition, filter ordering, and cutoff parameters right for the specific material class and synthesis route. It also requires an honest accounting of what failed in each screening campaign, so those failures inform the next iteration. The platform page describes how we structure the output of each screening run to make this kind of iteration possible. If you want to walk through how this would apply to your specific material class, get in touch and we can work through the parameters together.