← Case Studies | Catalyst Design

CASE STUDY — INVERSE DISCOVERY

Inverse Catalyst Discovery

This study evaluated a more stringent use case than simple catalyst scoring: define the catalyst brief, exclude established incumbent classes when needed, and determine whether the same full-stack workflow still converges to catalyst families that are consistent with the historical catalysis literature. Typical constrained inverse-search runs complete in seconds on local hardware; the public benchmark scored 1,000 full-stack API calls in 19.7 seconds.

Full API workflow
FLUX-enriched properties
Industrial catalyst convergence
Exclusion-mode discovery
Open benchmark Download benchmark JSON
9 / 9
inverse-search convergences
12 / 12
ranking tests passed
10 / 10
scenario alignment
19.7 s
1,000 full-stack API calls
96
FLUX-enriched references
What Was Tested

Preset rediscovery, scenario ranking, and exclusion-mode search

The study covered industrial ammonia synthesis, Fischer-Tropsch, ethylene oxide, hydrodesulfurization, and steam reforming, then constrained the search away from established incumbent classes to evaluate whether the outputs remained chemically plausible.

What FluxMateria Found

The workflow did more than rank known formulas

It reproduced promoted-iron, Ag/Cs/Re, Co/Mn, and CoMo or NiW families, then shifted into adjacent literature-supported candidate classes when incumbents were deliberately removed.

Why This Matters

A useful catalyst engine should remain stable under a changed design brief. It should move into plausible neighboring chemistry with interpretable tradeoffs rather than produce arbitrary compositions when constraints become stricter.

How FluxMateria got here

The value comes from the integration: ranking, support effects, tradeoffs, and inverse search all run inside one workflow through the live API path.

1

Define the catalyst brief

Set reaction family, design goal, and the practical constraints that matter for activity, selectivity, cost, stability, or scale-up.

Constraint spec
2

Generate or screen candidates

Search known candidates or enumerate metal, support, and promoter combinations from templates depending on the discovery question.

Candidate pool
3

Enrich with FLUX properties

The API path adds FLUX material properties and current surface-aware work-function signals before the catalyst layer scores the composition.

Enriched properties
4

Rank and explain tradeoffs

The engine scores activity, selectivity, thermal stability, poison resistance, regenerability, cost, and deactivation-linked readiness instead of returning a bare rank list.

Decision shortlist
5

Probe alternative candidate classes

Exclude incumbent elements or formula families and rerun. The key question is whether the ranking remains chemically plausible under explicit novelty constraints.

Alternative class

Literature Convergence

The first pass asked a baseline question: does FluxMateria converge to the same catalyst families that are already well established in the field?

Ammonia synthesis

FeKAlO4, FeBaCeO4, FeBaO
Ertl 2008, Schlögl 2003, Haber-Bosch promotion logic

The engine independently lands on promoted iron oxides, the same family uncovered through enormous experimental search for industrial ammonia synthesis.

Fischer-Tropsch SAF

CoMnSiO2, CoMnAlO3, FeMnAlO3
Dry 2002, de Klerk 2011, Morales 2005

Mn-promoted cobalt on inert supports is consistent with where Fischer-Tropsch selectivity optimization campaigns converged for high-value product distributions.

Ethylene oxide

AgCsReAlO3, AgCsClAlO3, AgCsAlO3
van Santen and Kuipers 1987, Grant and Lambert 2005

The top lane matches the modern industrial EO catalyst family: silver on alumina with alkali and rhenium or chloride promotion.

Hydrodesulfurization

CoMoAlS3, NiWAlS3, CoMoS3
Topsøe, Clausen and Massoth 1996

The engine rediscovers the promoted sulfide-on-support families that define industrial HDS, which is an appropriate baseline for a catalyst screening layer.

Steam reforming, coke-aware

RhAlO3, PtAlO3, RhCeO2
Rostrup-Nielsen and Hansen 1993

The engine captures the Rh > Pt > Ni coke-resistance logic and the linked activity-versus-cost tradeoff that shapes real reforming decisions.

Benchmark bridge

12 / 12 ranking tests, 10 / 10 scenario alignment
Same public scoring path

The case study is backed by the benchmark, not separated from it. The same layer now clears the corrected FT, ammonia-support, and WGS ordering cases through the deployed production stack.

Second-pass Exclusion Challenge

A more discriminating test followed rediscovery: remove established incumbent classes and determine whether the search remains inside chemically plausible catalysis.

Alternative class

Ammonia without ruthenium

WCeO2, WLaO2, WZrO2

Once Ru and Os are excluded, the engine shifts into Mo- and W-centered rare-earth support chemistry rather than toward arbitrary compositions. That is an appropriate adjacent search space for non-precious ammonia catalysis.

Literature context: Jacobsen 2001; Kojima and Aika 2001; Kitano 2012.
Alternative class

Fischer-Tropsch without cobalt

CuZnCeO3, CuZnAlO4

With cobalt removed, the engine moves into Cu-Zn-Ce chemistry rather than into chemically implausible outputs. The result redirects the search toward adjacent hydrogenation and syngas-conversion territory that the literature already treats as mechanistically credible.

Literature context: Davis 2009; Wei et al. 2017.
Alternative class

Ethylene oxide without silver

CuZnCeO3, CuZnAlO4, RuO2

Non-silver EO remains challenging experimentally, but the engine still moves into the same broad Cu-based and Ru-based research territory that selective-oxidation studies have already considered worth investigating.

Literature context: Linic and Barteau 2004; Torres et al. 2007.

Why this matters: not every exclusion-mode candidate is already industrially established, nor should it be. What matters is that the search remains chemically coherent and literature-adjacent when the incumbent answer is removed.

Benchmark Snapshot

The case study stands on top of the benchmark, not beside it.

Benchmark Backstop

The same public scoring path now clears the benchmark-relevant catalyst cases

The publication benchmark is full-stack and production-routed. That means the same enriched scoring path is what clears the corrected FT activity and support ranking cases, preserves Ru-support logic, preserves WGS ordering, and powers the inverse-search presets used in this case study.

12 / 12
ranking tests passed
Including FT activity, FT support, ammonia support, and WGS.
10 / 10
scenario alignment
Different catalyst briefs lead to different physically sensible top-ranked candidates.
98.4%
pairwise accuracy
The ranking engine gets the ordering right for nearly all public benchmark pairs.
51 / s
full-stack throughput
Production scoring path with enriched properties and surface-aware catalyst scoring.

Why this is novel, valuable, and faster

The result matters not because FluxMateria reproduced a single known catalyst, but because the same integrated workflow handled rediscovery, exclusion stress-testing, ranking, and shortlisting in a way that is typically split across separate tools and slower research loops.

Novel

The outputs were not seeded with target answers

FluxMateria was given reaction classes, constraints, and exclusions, not a pre-loaded table of industrial winners. The value is that the system still converged to promoted Fe, Ag/Cs/Re, Co/Mn, and sulfide-on-support families through the same public API path.

Valuable

The output is a decision-ready shortlist

The workflow compresses a broad catalyst space into a shortlist with support effects, tradeoffs, literature convergence, and adjacent alternative classes already exposed. That makes the next synthesis and reactor program more targeted.

Faster

Iteration is rapid enough to be practical

Small preset and exclusion searches resolve in seconds locally, and the hardened public benchmark sustained 51 candidates per second across 1,000 full-stack API calls. That is fast enough to rerun briefs, exclusions, and tradeoff assumptions instead of treating them as single-pass decisions.

What Remains

The remaining step is catalyst synthesis, reactor testing, and deactivation validation.

That is the point of the study: not to replace experimental catalysis, but to make the experimental program more targeted, more cost-efficient, and more technically defensible by forwarding a smaller set of higher-priority candidates and exclusion-tested alternative classes.

What happens next

The case study has done its job once the experimental program is narrower, more explicit, and better aligned to available laboratory resources.

1. Validate the shortlist

Synthesize the converged families first

Start with the rediscovered industrial lanes and confirm activity, selectivity, and stability under the intended operating window. That checks whether the same rank logic survives real catalyst preparation and reactor conditions.

2. Evaluate the alternative classes

Test the exclusion-pass candidates where incumbents fail on cost or supply

W- or Mo-supported ammonia classes and Cu-Zn-Ce exclusion results are not a claim of immediate replacement. They are priority adjacent experiments because the workflow already showed they are coherent enough to justify laboratory time.

3. Close the loop

Feed kinetics and deactivation data back into the layer

The next upgrade path is tighter experimental calibration: measured turnover, poison tolerance, sintering, coking, and support interaction data can turn this from strong screening into a more exact reactor-program planning tool.

Artifacts

The public-safe benchmark and the case study now live in the same website system and point to the same validated story.

Catalyst benchmark
Full public benchmark page for the API-only catalyst scoring and inverse-search workflow.
Open benchmark
Benchmark summary JSON
Public-safe summary artifact generated from the successful full-stack benchmark run.
Download JSON
Materials module
See how the catalyst layer sits inside the wider FluxMateria materials stack.
Open module page

Catalyst discovery should remain coherent under stricter constraints

Matching one known catalyst family is not sufficient. The stronger result is that FluxMateria also remains coherent when incumbent classes are removed and the search has to move into adjacent chemistry.

Benchmark details Back to Materials module