Computational searches for new battery electrodes generate hundreds of thousands of candidate crystals and then discard almost all of them. One recent solid-state battery pipeline started from 527,365 generated structures and kept 29, and along the way its generator produced crystals that were symmetric and apparently stable but chemically absurd, such as lithium in square-planar coordination (Generative Pipeline for Discovering Solid-State Battery Materials with Universal Atomistic Potentials). Much of the chemistry that would reject such candidates early is qualitative and chained. Selenium substitution expands the lattice, and lattice expansion speeds sodium diffusion. An insulating oxide needs a conductive carbon or core phase. Large volume swings during cycling demand a buffering matrix. Written as a formal ontology, these rules pass through many intermediate concepts that a screening script cannot observe. A screener sees composition, oxidation states, space group, local coordination and a computed bandgap. It does not see "lattice expansion" or "electron highway."
We propose to close this gap with logical forgetting, also called uniform interpolation: a family of algorithms that delete chosen vocabulary from an ontology while provably keeping every consequence expressible in the vocabulary that remains. We would write an ontology of conductivity, diffusion and strain constraints for transition-metal oxide, sulfide and phosphide electrodes. We would then forget everything except the vocabulary the screening pipeline can actually compute, and run the result as a compact, reasoner-free filter over generated and database candidates. The central point is that in screening, the target vocabulary is not a matter of user taste. It is fixed by what the measuring pipeline can compute. That removes the main practical objection to forgetting, the difficulty of choosing which terms to keep, and turns the method into a knowledge compiler with a completeness guarantee.
Core Claim
Projecting a mechanistic electrode ontology onto the vocabulary a pipeline can observe gives a filter with three properties together. First, it misses no constraint that the ontology implies over the observables. Second, it is small enough to run as vectorized checks over millions of structures. Third, it updates automatically when the ontology gains a new mechanism. Hand-written filters can match the first two properties for rules someone thought to write down. They cannot guarantee the observable constraints that emerge only by chaining several mechanisms. We will measure how many such emergent constraints exist in a realistic electrode ontology and whether they actually reject bad candidates. If there are few of them, the method has little to offer, and the experiments are designed so that this outcome would show up.
Relation to Prior Work
The forgetting machinery comes from description logic research. A description logic is a decidable fragment of first-order logic used for OWL ontologies. Forgetting Concept and Role Symbols in ALCOIHµ(∇, ⊓)-Ontologies eliminates both concept and role symbols with an Ackermann-style substitution calculus. Forgetting and Unfolding for Existential Rules does the same for rule languages by unfolding rules, and it shows that deleting intermediate predicates syntactically destroys query answers. Efficient Computation of General Modules for ALC Ontologies computes "general modules," which are not restricted to subsets of the original axioms, keep axioms short, and bound size to single-exponential. Efficient Extraction of EL-Ontology Deductive Modules extracts axiom-subset modules through a SAT encoding. These papers supply our method. We reuse their tools rather than invent a calculus.
All of them were evaluated on biomedical or generic corpora such as SNOMED CT, NCI and ORE 2015, where the kept vocabulary is chosen by a user or at random. Metadata-based Term Selection for Modularization and Uniform Interpolation of OWL Ontologies treats the choice of that vocabulary as an open bottleneck and ranks candidate terms with embeddings. In screening, the bottleneck largely disappears, because the vocabulary is fixed by the descriptors a pipeline can compute.
Two strands of work motivate the proposal by documenting the gap. Towards Ontology Reshaping for KG Generation with User-in-the-Loop: Applied to Bosch Welding flattens a manufacturing ontology heuristically and explicitly leaves uniform interpolation to future work. The materials ontologies model static structure and provenance, in the case of An Ontology for the Materials Design Domain, or offer untyped facet taxonomies in which "Performance" resists formalization, in the case of Materials Science Ontology Design with an Analytico-Synthetic Facet Analysis Framework. Neither is axiomatized for reasoning about conduction.
Two further works set the bar we must beat, and they also limit what can be claimed. LLEMA: Evolutionary Search with LLMs for Multi-Objective Materials Discovery injects chemistry rules as prompt text and is strong empirically. A Generative Model for Inorganic Materials Design, MatterGen, is our candidate generator. It has a known bias toward low-symmetry P1 output, which gives a filter something to catch.
Research Program
The first component is a constraint ontology for conversion and alloying electrodes of transition metals. Its axioms come from the mechanisms reported in the electrode literature in the file:
- carbon hybridization and chemical bonding to graphene (Song et al., Q. Wang et al.);
- nitrogen doping (Guo et al.);
- anion substitution that expands the lattice (K. Li et al.);
- conductive-core heterostructures (Construction of CuCo2O4@NiFe-LDH Core–Shell Heterostructure for High-Performance Hybrid Supercapacitors).
It also includes standard crystal-chemistry plausibility rules: charge balance, cation–coordination compatibility, and the ban on square-planar lithium that motivated the solid-state battery pipeline's failures. We target the description logic ALC, the basic logic with negation, conjunction, disjunction and role restrictions. If disjunction proves unnecessary we fall back to the simpler logic EL. Numeric quantities such as bandgap, predicted volume change on sodiation and energy above the convex hull are discretized into observable concept names like BandgapAbove3eV. This is a deliberate loss of precision, and we will report it.
The second component fixes the observable signature Σ_obs, meaning the set of concept and role names that pymatgen-style analysis and MACE-level relaxation (MACE) can assign to a crystal. We then compute the uniform interpolant of the ontology over Σ_obs with an Ackermann-based forgetter. Finite interpolants need not exist in ALC. When they do not, or when they blow up in size, we switch to a GEMO general module, which is the reason that tool is in the program. The output is translated into vectorized predicate checks over a candidate table, and each rejection is traced back to the mediating concepts in the full ontology so that a chemist can read why a candidate was rejected.
The third component is an update protocol. New mechanism axioms are added in batches, the interpolant is recompiled, and the diff over Σ_obs shows exactly which new observable constraints each mechanism implies.
Baselines
The naive reference point has two parts. One is no filter: every generated candidate goes straight to MACE relaxation. The other is syntactic deletion: we drop every axiom that mentions a non-observable symbol and run whatever remains. Both use the same candidate pools and the same relaxation budget. Syntactic deletion is the fair "cheap formal" baseline, because Forgetting and Unfolding for Existential Rules predicts that it loses entailments, and we want to measure how much that loss costs in a real screening setting.
The strongest practical competitor is an expert-written filter. A materials scientist who has not seen the interpolant gets the same mechanism papers and the same number of hours we spend writing the ontology, and writes the observable-level checks directly in code. Most groups would do this. It has no reasoning overhead, and it will probably capture every rule stated explicitly in a paper. Beating it means rejecting significantly more candidates later confirmed implausible or unstable, at equal or lower false rejection of known stable materials, with the gap traceable to constraints that arise only by chaining.
The strongest formal competitor runs a description logic reasoner on the full ontology for each candidate, treating the candidate as a set of facts. A variant first shrinks the ontology to a ForMod deductive module. These give the same answers as the interpolant by construction, so the comparison is throughput and portability on identical hardware. The reported reference is that ForMod runs on average 99.79 times faster than the Zoom algorithm on SNOMED CT (Efficient Extraction of EL-Ontology Deductive Modules). No reported result was found on Lacuna for per-candidate reasoning throughput over crystal databases.
The strongest discovery-loop competitor is a LLEMA-style critic, in which a language model receives the same mechanism rules as text and accepts or rejects candidates. On LLEMABench's wide-bandgap task, LLEMA reports a hit rate of about 33.6%, against 6.5% for MatterGen (LLEMA). No reported result was found on Lacuna for any method on transition-metal conversion-electrode screening. We would therefore report our own numbers for all systems on a shared task, and also run LLEMABench's wide-bandgap task to anchor ours against theirs.
Two ablations isolate the key ingredient. In the first, we keep the mediating concepts rather than forgetting them, and check candidates only against explicitly stated observable rules. The difference between this and the full interpolant is exactly the contribution of the implied, chained constraints. In the second, we compute the interpolant over an embedding-chosen signature from NN-RANK instead of the pipeline-defined Σ_obs. This tests whether fixing the signature by what can be measured matters.
Why not just…?
Why not just have a chemist hand-code the filter? This is the real threat. If the electrode ontology contains few rules that emerge only by chaining, the hand-written filter will match ours and the proposal reduces to a formal certificate of something cheap. We do not argue around this. The chaining ablation and the expert baseline measure it directly, counting interpolant constraints absent from the expert filter and how many real candidates those constraints reject. A second, independent test is the update experiment: after each new mechanism batch, we ask whether the expert revises the filter as completely as recompiling the interpolant does.
Why not just run a reasoner on the full ontology? A reasoner gives the same verdicts, so the question is operational. If per-candidate reasoning over 10^5 to 10^6 structures fits the budget, the interpolant's advantage shrinks to three things: it runs as plain dataframe or SQL checks inside generation loops, it can serve as a rejection step inside MatterGen sampling without a reasoner in the loop, and it gives a readable diff when knowledge changes. We will report wall-clock time for both and say plainly if reasoning is fast enough.
Why not just prompt a language model with the rules, as LLEMA does? Prompted rules carry no completeness guarantee, and the error memory in LLEMA is statistical. Still, LLEMA's hit rates are strong, and language models may apply implicit chemistry that our ontology lacks. We will compare rejection precision and recall on the same pools. We will also test the combination, with our filter used as a hard gate in front of a LLEMA-style loop. If the combination does not beat either method alone, that is informative.
Why not just extract a module instead of forgetting? Modules such as those in Efficient Extraction of EL-Ontology Deductive Modules are subsets of the original axioms, so they keep the mediating vocabulary and still need a reasoner. The full-reasoner baseline with a module covers this alternative. General modules sit in between, and we already use them as the fallback when finite interpolants fail.
Proposed Experiments
The first check is a consequence audit, which is decisive for correctness. For every pair of atomic observable concepts, and for a few thousand sampled compound observable queries, we compare what the interpolant entails with what the full ontology entails, using a standard reasoner. We require exact agreement, and we report the fraction of entailments lost by syntactic deletion. We also run the forgetter on the ORE 2015 ALC corpus as a sanity check. Axiom length should be close to GEMO's reported median longest axiom of 3 (Efficient Computation of General Modules for ALC Ontologies).
The second check measures filter quality on labelled pools. Positives are transition-metal oxides, sulfides and phosphides from Materials Project, de-duplicated via LeMat-Bulk (LeMat-Bulk). Negatives come from two sources. One is MatterGen samples, generated with the public code (microsoft/mattergen), that fail MACE relaxation or sit far above the hull. The other is controlled corruptions such as square-planar lithium or broken charge balance. The metrics are rejection recall on negatives, false rejection on positives, and throughput. All baselines and ablations run on the same pools.
The third check is a funnel test. On a fixed budget of MatterGen candidates in electrode-relevant chemistries, we compare the fraction of filter-passing structures that survive MACE relaxation with and without each filter. The reference is the 527,365-to-29 funnel reported by the solid-state battery pipeline (Generative Pipeline for Discovering Solid-State Battery Materials with Universal Atomistic Potentials), though that funnel is for a different target, so it serves only as context.
The fourth check is the update experiment. We add mechanism batches, for example heterostructure axioms after carbon-hybridization axioms, recompile, and measure how many new observable constraints appear and how many the expert filter picked up.
The proposal succeeds if three things hold. The audit must show exact agreement. The chaining ablation must reject a non-trivial share of MACE-confirmed failures that the expert filter lets through. And throughput must beat per-candidate reasoning by at least an order of magnitude. If the chaining contribution is negligible, we will publish that as a negative result on how much knowledge is hidden in intermediate concepts of materials ontologies.
References
- Generative Pipeline for Discovering Solid-State Battery Materials with Universal Atomistic Potentials
- Forgetting Concept and Role Symbols in ALCOIHµ(∇, ⊓)-Ontologies
- Forgetting and Unfolding for Existential Rules
- Efficient Computation of General Modules for ALC Ontologies
- Efficient Extraction of EL-Ontology Deductive Modules
- Metadata-based Term Selection for Modularization and Uniform Interpolation of OWL Ontologies
- Towards Ontology Reshaping for KG Generation with User-in-the-Loop: Applied to Bosch Welding
- An Ontology for the Materials Design Domain
- Materials Science Ontology Design with an Analytico-Synthetic Facet Analysis Framework
- LLEMA: Evolutionary Search with LLMs for Multi-Objective Materials Discovery
- A generative model for inorganic materials design
- Chemically Bonded Phosphorus/Graphene Hybrid as a High Performance Anode for Sodium-Ion Batteries
- Reduced Graphene Oxide-Wrapped FeS2 Composite as Anode for High-Performance Sodium-Ion Batteries
- N-Doped Modified Graphene/Fe2O3 Nanocomposites as High-Performance Anode Material for Sodium Ion Storage
- S, Se-Codoped Dual Carbon Coating and Se Substitution in Co-Alkoxide-Derived CoS2 Through SeS2 Triggered Selenization for High-Performance Sodium-Ion Batteries
- Construction of CuCo2O4@NiFe-LDH Core–Shell Heterostructure for High-Performance Hybrid Supercapacitors
- MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields
- LeMat-Bulk: aggregating, and de-duplicating quantum chemistry materials databases