Back to blog
Technical

How We Design an Inorganic Screening Pipeline: Three Decisions That Matter

How We Design an Inorganic Screening Pipeline: Three Decisions That Matter

Building an inorganic screening pipeline feels, from the outside, like it should be mostly an engineering problem. You have a candidate library, you have a scoring model, you rank and filter. In practice the design decisions that matter most are not about the model or the compute: they are about how you frame the search problem, how you order the filters, and what rules you use when candidates are effectively tied. Each of those three decisions shapes the synthesis queue in ways that are hard to back out after the fact.

This is not a comprehensive treatment of pipeline architecture. It is a description of the three design choices where we have seen the most consequential variation, and where getting it wrong costs the most in downstream synthesis time.

Decision 1: Candidate Generation Strategy

The way you generate the initial candidate pool determines the ceiling on what the screen can discover. Two broad approaches exist, and they answer different questions.

Substitution-based generation

You take a set of known prototype structures (layered oxides, spinels, perovskites, rocksalts, and so on) and systematically substitute elements at each crystallographic site using a substitution matrix derived from chemical similarity, ionic radius tolerance, and oxidation state compatibility. The result is a large but chemically grounded library of hypothetical compounds that are structurally plausible by construction. Most high-throughput screening work in battery cathodes and oxide ceramics uses this approach.

The limitation is that you are constrained to the prototype topologies you started with. You will not discover a novel structure type that has no known analogue in your prototype library. If your screening goal is to find the best composition within a known structural family, this is fine. If you are looking for structurally novel candidates, it is not.

Generative enumeration

You enumerate candidate structures from crystal symmetry space groups and stoichiometric constraints, without requiring a specific prototype. This produces a larger and more diverse candidate space, including structure types with no known experimental realization. The tradeoff is that a much larger fraction of generated candidates will be unphysical: overlapping atoms, implausible coordination geometries, or steric problems that make synthesis practically impossible regardless of what the energy model says.

Generative enumeration requires an early-stage geometric feasibility filter before you run any property evaluation, or you waste expensive model calls on geometrically degenerate structures. The filter does not need to be exact: it needs to be fast and have a low false-negative rate (do not reject synthesizable structures).

We use substitution-based generation as the default for most materials classes we work with, because the typical customer problem is "find better compositions within this family" rather than "discover an entirely new structure type." For exploratory work where structure novelty is the goal, we switch to generative enumeration with a geometry pre-screen.

Decision 2: Filter Ordering

Once you have a candidate library, you need to apply multiple filters before synthesis dispatch: thermodynamic stability, target-property thresholds, synthesis feasibility, and possibly additional domain-specific criteria. The order in which you apply those filters has a significant effect on pipeline cost and on what survives to the shortlist.

Why ordering matters

Each filter has a cost (compute time per structure) and a pass rate (the fraction of structures that survive). The cheapest filter to run is almost always the thermodynamic stability check: convex hull distance computation against a reference phase diagram is fast, and for most inorganic composition spaces, a meaningful fraction of hypothetical candidates (often 40 to 70 percent, depending on how broadly you defined the candidate space) sit more than 100 meV/atom above the hull. Running your expensive target-property model on those structures is waste.

The general rule is: run filters in order of increasing cost and apply the most aggressive filters first, subject to one constraint. If a later filter is both cheap and highly selective, it may be worth running it before a more expensive early filter even if cost-rank would place it later. The sequence we use for most inorganic screens is:

  1. Geometric feasibility (fast, catches degenerate structures from generative enumeration)
  2. Charge balance and oxidation state compatibility (algebraic, near-zero cost)
  3. Thermodynamic stability via convex hull distance (MLIP energy evaluation, milliseconds per structure)
  4. Target-property threshold (MLIP property prediction, also fast but model-dependent)
  5. Synthesis feasibility score (heuristic function over predicted bond lengths, coordination numbers, precursor availability)

Filters 1 and 2 together typically reduce the candidate pool by 20 to 40 percent at near-zero cost. Filter 3 removes another 40 to 70 percent. By the time you reach filter 4, you are running the property model on 10 to 30 percent of the original candidate pool rather than the full library. This matters when you are screening tens of thousands of structures and property model evaluation has a meaningful cost per call.

The false-negative problem

Every filter you add is also a potential source of false negatives: good candidates that get rejected because your stability or feasibility thresholds are set too tight. We are not suggesting that tighter filters are always wrong; we are noting that filter threshold calibration deserves the same attention as model accuracy. A thermodynamic stability filter set at 50 meV/atom above hull will reject a different and smaller set of candidates than one set at 100 meV/atom. Some materials teams want tight filters to reduce synthesis volume. Others want loose filters because the cost of a false negative (a missed candidate) exceeds the cost of one extra synthesis run. That is a business and cost question, not a modeling question, and the pipeline designer should get an explicit answer to it before fixing thresholds.

Decision 3: Tie-Breaking Rules

After filtering, you will have a ranked shortlist. The ranking function itself, typically a combination of convex hull distance and target-property score, produces a clean ordering for the top and bottom of the list. But in the middle of the distribution, particularly when the score differences between candidates are smaller than your model's prediction error, the ranking is effectively a tie. How you break those ties determines which candidates the synthesizer actually sees.

What makes a good tie-breaker

Common approaches include: preferring candidates with simpler precursor chemistry (fewer elements, precursors that are commercially available and are known to be compatible), preferring candidates that are structurally similar to compounds that have been successfully synthesized in the past (either in your internal data or in the literature), and preferring candidates where the predicted properties are stable under small perturbations of the structure (indicating a deep basin in the energy landscape rather than a fragile local minimum).

Synthesis-friendliness as a tie-breaker is practical but has a real cost: it biases the shortlist toward known chemistry and reduces the probability of discovering genuinely novel compounds. This is worth stating explicitly to the materials team before implementing. If the goal is to find the best material within a known accessible space, synthesis-friendliness is the right tie-breaker. If the goal is exploration, it is the wrong one.

Diversity as a tie-breaking criterion

A third approach is to use structural and compositional diversity as a soft tie-breaker: among candidates with similar scores, prefer candidates that are structurally different from each other. This increases the information return per synthesis run, since synthesizing five structurally similar candidates teaches you less than synthesizing five candidates that sample different regions of the space. We use a diversity term in our synthesis queue when the candidate pool size and synthesis capacity budget allow for it.

What the Synthesizer Actually Sees

All three of these decisions compound. A substitution-based library with a tight thermodynamic filter and synthesis-friendliness tie-breaking will produce a shortlist that is dense around known compound families, has low synthesis risk, and has a high probability of producing phases but a lower probability of producing genuinely novel ones. A generative enumeration library with looser filters and diversity tie-breaking will produce a shortlist that is more exploratory, has higher synthesis uncertainty, and will fail more often per run while producing more information per failed run.

There is no universally correct pipeline design. The right design depends on where the materials team is in its discovery cycle, how much synthesis throughput they have available, and what the actual value of a novel compound discovery is relative to an incremental improvement in a known family. What we try to do at AstraIQ is make those three decisions explicit before we configure the pipeline, rather than treating them as implementation details. They are not implementation details. They are the decisions that determine what the screening system actually optimizes for.

If you are working on a screening problem and want to walk through how these choices map onto your specific material class and synthesis capacity, get in touch through the contact page. We find that the most useful early conversation is not about model accuracy but about candidate generation scope and synthesis budget.