AI materials-discovery engine (candidate generation and in-house synthesis)

What the technology is

The workflow consists of three linked parts:

1. Candidate generation — a machine-learning model proposes materials, crystal structures, or molecules likely to have desired properties.
2. Scoring/ranking — physics-based methods such as density functional theory (DFT), molecular dynamics, or model-predicted stability/activity scores rank candidates.
3. In-house synthesis and validation — an autonomous or semi-autonomous laboratory synthesizes the candidate, measures the property, and feeds the result back into the model.

The key claim of the modern AI-materials stack is not that AI alone can replace physical chemistry; it is that a closed loop shrinks the search space and accelerates the cycle between prediction, experiment, and re-training.

The strongest peer-reviewed evidence for the candidate-generation half is Google DeepMind’s GNoME paper. The strongest peer-reviewed evidence for the in-house synthesis half is the University of Liverpool mobile robotic chemist and MIT’s AI-informed flow-synthesis platform.

Source assessment

Tier 1 sources used

• Peer-reviewed paper, Nature: Merchant et al., “Scaling deep learning for materials discovery,” Nature 624, 80–85 (2023), DOI: 10.1038/s41586-023-06735-9.

URL: nature.com

• Peer-reviewed paper, Nature: Burger et al., “A mobile robotic chemist,” Nature 583, 237–241 (2020), DOI: 10.1038/s41586-020-2442-2.

URL: nature.com

• Peer-reviewed paper, Science: Coley et al., “A robotic platform for flow synthesis of organic compounds informed by AI planning,” Science 365, eaax1566 (2019), DOI: 10.1126/science.aax1566.

URL: science.org

• Peer-reviewed review, Nature: Butler et al., “Machine learning for molecular and materials science,” Nature 559, 547–555 (2018), DOI: 10.1038/s41586-018-0337-2.

URL: nature.com

• Granted patent, USPTO: U.S. Patent No. 10,438,096 B2, “Systems and methods for applying a convolutional neural network to identify molecules that bind to a target protein,” Atomwise Inc., granted 2019-10-08.

URL: patents.google.com

Tier 2 sources used

These are company technical documents or self-reported engineering blogs. They are not Tier 1 for performance claims, but they are useful for adoption/deployment context.

• Google DeepMind blog: “Millions of new materials discovered with deep learning,” Google DeepMind, 2023-11-29.

URL: deepmind.google

• Microsoft Learn: “What is Azure Quantum Elements?” Microsoft, accessed 2026-04-26.

URL: learn.microsoft.com

• Schrödinger materials science page: Schrödinger, Inc., materials science platform overview.

URL: schrodinger.com

Securities filings used

• Alphabet Inc., Form 10-K for FY2023, SEC EDGAR, filed 2024-01-30.

URL: sec.gov

• Schrödinger, Inc., Form 10-K filings list, SEC EDGAR, CIK 0001490978.

URL: sec.gov

• Microsoft Corp., Form 10-K filings list, SEC EDGAR, CIK 0000789019.

URL: sec.gov

• Recursion Pharmaceuticals, Inc., Form 10-K filings list, SEC EDGAR, CIK 0001601830.

URL: sec.gov

No arXiv preprints are used in the load-bearing part of this analysis because journal versions exist for the main papers.

Candidate generation: what the peer-reviewed evidence actually says

3.1 GNoME’s reported output

The GNoME paper reports the following load-bearing numbers:

• 2.2 million new crystal structures were generated by the model.
• 381,000 of those structures are predicted to be stable, meaning the authors consider them likely enough to be candidates for experimental realization.
• 736 structures were reported as independently experimentally realized.

Citation: Merchant et al., Nature 624, 80–85 (2023), DOI: 10.1038/s41586-023-06735-9.

This is a major distinction. “2.2 million structures” is not “2.2 million synthesized materials.” The 381,000 stable candidates are model-predicted, not synthesized. The 736 experimentally realized structures are the narrow external-validity slice. That gap is the central commercialization risk in candidate generation.

[SOURCE-BOUND: GNoME paper] The 2.2 million/381,000/736 numbers are bound to the DeepMind-led Nature paper. The paper is peer-reviewed, so it satisfies Tier 1 for the claim “the model reported this output,” but it is not independent proof that all 381,000 candidates are synthesizable or commercially useful. [SINGLE-SOURCE] for the model-output claim.

3.2 Method statement

The GNoME model is a graph neural network trained on crystal-structure data from the Materials Project and other sources. The paper reports that the model can traverse the space of possible inorganic crystals more efficiently than brute-force high-throughput screening.

Why this matters for the thesis: the economic value of exact DFT screening is constrained by compute cost. A learned model that replaces or filters DFT candidates could reduce compute per candidate. However, the paper does not provide a clean USD-per-candidate cost figure, and any such figure would require primary source validation from the authors or a third-party cost model.

3.3 Marketing versus peer-reviewed claim

Google DeepMind’s blog title says “Millions of new materials discovered with deep learning.” The peer-reviewed abstract is more careful: it distinguishes generated structures from stable candidates and from experimentally realized materials.

This is a source-tier contradiction:

• Marketing/Tier 2: “Millions of new materials discovered.”
• Peer-reviewed/Tier 1: 2.2M structures; 381,000 stable candidates; 736 independently experimentally realized.

The neutral reading is that the word “discovered” in the blog means “model-generated,” not “synthesized.” Investors should track the experimental validation count because that is the number that matters for physical-world feasibility.

In-house synthesis: the bottleneck evidence

4.1 Mobile robotic chemist

The Liverpool robotic chemist paper is strong Tier 1 evidence that autonomous in-house synthesis can operate as a real experimental loop.

Load-bearing numbers from the abstract:

• The robot ran for 8 days.
• It performed 688 experiments in that period.
• It used Bayesian optimization over a 10-variable experimental space.
• It discovered a photocatalyst mixture reported to be six times more active than the initial formulation.

Citation: Burger et al., Nature 583, 237–241 (2020), DOI: 10.1038/s41586-020-2442-2.

[SINGLE-SOURCE] for the specific 688/8-day throughput. This is one lab, one robot, one chemistry problem. It is an existence proof, not a broad reliability guarantee.

4.2 AI-informed flow synthesis

The MIT/Science paper demonstrates a different model: an AI planning system creates synthetic routes, and a robotic flow platform executes them.

Citation: Coley et al., Science 365, eaax1566 (2019), DOI: 10.1126/science.aax1566.

This matters because it shifts synthesis from batch-style “make one material in a flask” toward programmable, reproducible flow chemistry. The technical claim relevant here is not that this specific 2019 platform can make every GNoME candidate; it is that synthesis can be partially automated through a planning-execution pipeline.

4.3 Why synthesis is the bottleneck

GNoME generates hundreds of thousands of stable candidates. The autonomous chemistry evidence suggests throughput on the order of tens to low hundreds of experiments per day per setup. That asymmetry is the core reason the phrase “candidate generation and in-house synthesis” must be treated as a two-part system, not a solved end-to-end product.

Patent analysis: what the granted patent does and does not prove

The best retrievable granted patent for this note is:

U.S. Patent No. 10,438,096 B2

Title: “Systems and methods for applying a convolutional neural network to identify molecules that bind to a target protein”

Assignee: Atomwise Inc.

Grant date: 2019-10-08

URL: patents.google.com

This is a granted patent, not an application. That is the distinction the hard rules require: a granted patent supports an IP-moat claim more strongly than a pending application because the USPTO has completed examination and issued claims.

But there are limitations:

• The patent is in molecular binding prediction, not inorganic crystal discovery. It is therefore relevant to the broader “AI candidate generation” thesis, but it is not a patent on GNoME-type crystal generation.
• A granted patent is not proof of performance. The patent office does not verify that the neural network achieves any specified hit rate, selectivity, or economic return.
• The patent does not by itself create commercial viability. It is evidence of recorded claims, not evidence of revenue.

For GNoME-specific patent coverage, this note did not retrieve a granted patent solely covering GNoME’s graph-network crystal generation. That absence should not be filled with an unverified patent number. A search for a specific GNoME-assigned granted patent requires primary source validation.

Company filing cross-reference

6.1 Alphabet / Google DeepMind

Alphabet’s FY2023 Form 10-K does not break out “materials discovery” as a separate revenue line. GNoME is research output from Google DeepMind, not a disclosed standalone P&L segment.

Source: Alphabet Inc. Form 10-K for FY2023, SEC EDGAR, filed 2024-01-30.

URL: sec.gov

Discrepancy: the DeepMind blog markets “millions of new materials,” but the SEC filing does not quantify any commercial materials-discovery revenue. This is not necessarily an inconsistency; it means the product-level economics are not disclosed in the public filing.

6.2 Microsoft

Microsoft markets Azure Quantum Elements as a cloud service for chemistry and materials simulation. The technical documentation describes a workflow of compute, AI models, and quantum-ready infrastructure.

Source: Microsoft Learn, “What is Azure Quantum Elements?” accessed 2026-04-26.

URL: learn.microsoft.com

Microsoft’s Form 10-K does not appear to break out Azure Quantum Elements as a separate segment. The relevant SEC disclosure is at the consolidated Azure/cloud level, not the AI-materials product level.

Source: Microsoft Corp. Form 10-K filings list, SEC EDGAR.

URL: sec.gov

6.3 Schrödinger

Schrödinger sells computational chemistry and materials science software and also performs in-house drug discovery. Its materials science page explicitly markets a physics-based and machine-learning platform for materials design.

Source: Schrödinger materials science platform.

URL: schrodinger.com

SEC filings show Schrödinger is a public filer with software and discovery activities, but the retrievable filings do not isolate an “AI materials discovery engine” unit with a separate margin or revenue figure.

Source: Schrödinger, Inc. Form 10-K filings list, SEC EDGAR.

URL: sec.gov

Quantitative worked calculation

7.1 Throughput gap calculation

The question: if GNoME’s 381,000 predicted stable candidates were to be experimentally validated using one mobile robotic chemist at the throughput reported in Burger et al., how long would it take?

Formula:

\[

T_{\text{days}} = \frac{N_{\text{candidates}}}{R_{\text{experiments/day}}}

\]

Inputs:

• \(N_{\text{candidates}} = 381,000\) predicted stable candidates

Source: GNoME, Nature 624, 80–85 (2023), DOI: 10.1038/s41586-023-06735-9.

• \(R_{\text{experiments/day}} = 688 / 8 = 86\) experiments/day

Source: Burger et al., Nature 583, 237–241 (2020), DOI: 10.1038/s41586-020-2442-2.

Calculation:

\[

T_{\text{days}} = \frac{381,000}{86} = 4,430.2

\]

\[

T_{\text{years}} = \frac{4,430.2}{365} \approx 12.1

\]

So, under these assumptions, one reported autonomous-lab throughput would require roughly 12 years to validate every GNoME-predicted stable candidate.

7.2 Sensitivity

This calculation is [DERIVED] because it combines two numbers from two separate single-institution sources. It assumes:

• 100% of candidates are synthesizable;
• one setup runs continuously;
• no duplicate or replicate experiments;
• no reagent, cost, safety, or scheduling constraints;
• synthesis success is binary and immediate.

These assumptions are strong. The point is directional: synthesis throughput, not model candidate generation, is the likely bottleneck.

Parameter table

Contradictions and unresolved issues

9.1 Marketing “millions” vs experimental realization

DeepMind’s blog says “millions of new materials discovered.” The Nature paper reports 2.2M structures, 381,000 predicted stable candidates, and 736 independently experimentally realized candidates. If “discovered” means “experimentally realized,” the blog claim is an overstatement. If it means “model-generated,” it is technically compatible with the paper but still marketing language.

9.2 Patent scope vs product scope

The Atomwise granted patent is for CNN-based identification of molecules that bind to a target protein. That is not inorganic crystal generation. It is strong for the broad thesis that AI-based candidate generation can be patented, but it does not prove DeepMind or Microsoft has a granted patent on the GNoME workflow.

9.3 Filing gap

Public filings do not isolate revenue or margins for “AI materials discovery engines.” Therefore, any specific margin, ARR, or market-size claim requires primary source validation.

Unit economics / business model

Who monetizes

• Cloud providers monetize compute and model access. Microsoft’s Azure Quantum Elements is a consumption/credit-style cloud workflow; the exact per-job or per-hour pricing is not made public in the cited Microsoft Learn page and requires primary source validation.
• Software vendors monetize licenses/subscriptions for materials-simulation software. Schrödinger’s public materials-science page markets a platform, and its SEC filings show software/discovery operations, but materials-specific unit economics are not separately disclosed in the filings retrieved.
• Model owners monetize IP through patents, licenses, or internal R&D cost avoidance. The Atomwise patent is evidence of a patent asset, not evidence of royalty or licensing revenue.
• Lab operators monetize the physical synthesis loop by reducing cost per experiment or by selling validated materials/compounds. The Liverpool robotic chemist paper gives throughput evidence, but it does not report a USD cost per experiment.

Typical margins

No materials-discovery-specific margin figure was retrieved from the Tier 1 or SEC sources. Software gross margins in general are often high, but applying that generalization to Schrödinger’s materials engine without a filed breakdown would be unsupported. Therefore: materials-specific margins need primary source validation.

Who captures value

The likely value capture is split:

• Cloud/software layer: captures a per-seat or consumption fee.
• Patent layer: captures licensing/defensive value only if claims are enforced and relevant.
• Experimental layer: captures the compounding benefit of a proprietary training loop; the company that owns both the model and the in-house lab has the strongest feedback advantage.

Disconfirming Evidence

The thesis would be weakened or falsified by:

1. Failed independent synthesis campaigns — if multiple independent labs attempt high-ranked GNoME candidates and achieve a much lower synthesis/validation rate than the paper implies, the candidate-quality claim weakens.
2. Unreplicated autonomous throughput — if the 688-experiments-in-8-days result cannot be reproduced outside the originating lab, the throughput bottleneck cannot be solved by simply copying that robot.
3. Invalidated or narrowly construed patents — if Atomwise’s granted patent is invalidated, not enforced, or held irrelevant to materials candidate generation, the IP-moat argument weakens.
4. No disclosed commercial economics — if public companies continue to avoid disclosing AI-materials revenue or margins, the investment case remains dependent on marketing rather than filed evidence.
5. Chemistry cost inferiority — if the true cost per validated material remains higher than conventional high-throughput experimentation after accounting for failed syntheses and lab overhead, the AI-plus-automation loop would not be economically superior.