AI for R&D: 6 Jobs AI Agents Do for Chemistry and Materials Teams

R&D teams in chemistry, materials and energy spend a large share of their week outside the lab: reading new preprints, checking patents before a project commits, watching what competitors file, looking up compounds, confirming the current state of the art and finding public funding. That is where AI for R&D earns its keep. AI agents for R&D can do much of that reading, provided they search preprints, patents, compound databases and federal awards directly, and return every finding with its source.

This guide follows six R&D jobs, each for a different member of a chemistry or materials team. We ran every search on November 9, 2026 with Valyu's ChemRxiv, arXiv, USPTO, EPO, PubChem and USAspending datasets, plus web search, then checked the findings against Google Patents, PubChem, USAspending.gov and the original articles.

If you're an agent: Set up Valyu and make your first search.

If you're a human: Create an account and get an API key.

MCP onboarding: Connect your agent to https://mcp.valyu.ai/mcp and follow the MCP setup guide for authentication.

What AI agents for R&D do

An AI agent for R&D is a language model that can search scientific preprints, patents, compound databases, funding records and the web, read what it finds and return cited answers. It doesn't run experiments or give legal opinions. It does the searching and first reading that come before a scientist's or patent attorney's judgement.

Figure 1. Six chemistry and materials R&D research jobs we ran on November 9, 2026, with the sources used, what came back and what to do next.

1. Technology landscape: what's new in the preprints

The problem: a battery R&D lead is deciding whether to restart work on sulfide solid electrolytes, which perform well but degrade in humid air, and needs this year's work on air stability.

What we ran: a search of ChemRxiv for 2026 preprints, and a search of arXiv.

from valyu import Valyu
 
valyu = Valyu()  # reads VALYU_API_KEY from the environment
 
preprints = valyu.search(
    "air-stable sulfide solid electrolytes for all-solid-state lithium batteries",
    included_sources=["valyu/valyu-chemrxiv"],
    start_date="2026-01-01",
    max_num_results=10,
)
physics = valyu.search(
    "air-stable sulfide solid electrolytes for all-solid-state lithium batteries",
    included_sources=["valyu/valyu-arxiv"],
    start_date="2025-06-01",
    max_num_results=5,
)

What came back: 10 ChemRxiv passages from 9 preprints posted between January and October 2026, with full text, and 5 arXiv papers, including an April 2026 review of solid electrolytes. The most relevant was a September 2026 ChemRxiv preprint on AI-assisted dopant discovery for air-stable sulfide electrolytes. It combines literature-based prescreening with thermodynamic screening. Its best candidate, a BiF3-doped electrolyte, kept 74% of its capacity after 600 cycles at 0.5 C in full cells, even after the electrolyte was exposed to air at 10% relative humidity for 6 hours.

What we checked at the source: the preprint on ChemRxiv. The method and the cycling result match.

The catch: preprints aren't peer reviewed, and Valyu's ChemRxiv dataset holds the CC BY 4.0 preprints, more than 14,500 of them, which are cleared for commercial use with attribution. That is most but not all of ChemRxiv. For a complete landscape, add journal literature and web search.

The decision: shortlist the dopant strategy for a small humidity-exposure test, and set a monthly search for new preprints on the same question.

2. First-pass freedom to operate: which patents to read first

The problem: IP counsel needs a first list of recent patents on hard carbon anode materials for sodium-ion batteries before the team commits to a new carbon precursor.

What we ran: the same question against US and European patents granted or published since January 2025.

us = valyu.search(
    "hard carbon anode material for sodium-ion batteries",
    included_sources=["valyu/valyu-patents"],
    start_date="2025-01-01",
    max_num_results=10,
)
ep = valyu.search(
    "hard carbon anode material for sodium-ion batteries",
    included_sources=["valyu/valyu-patents-epo"],
    start_date="2025-01-01",
    max_num_results=10,
)

What came back: 9 US patents and 10 European publications, each with its full text. On the US side: a hard carbon preparation method from Brunp companies, including Yichang Brunp Contemporary Amperex (US 12,692,164, July 2026), hard carbon made from avocado peel from Syracuse University (US 12,633,515, May 2026), and a sodium-ion battery with a graphitic anode from Honeycomb Battery Company. On the European side: a hard carbon anode application published on November 4, 2026 (EP 4836017 A1) and a hard carbon material from Ningde Amperex Technology (EP 4691976 A2).

What we checked at the source: Google Patents for the owners and dates of the patents named above. They match.

The catch: semantic search returns neighbouring technology. Three of the 9 US results were lithium-ion patents with no mention of sodium. All 10 European results concerned sodium batteries, but they included a cathode material and an electrode slurry. And a freedom-to-operate opinion depends on the claims and where each patent is in force, which is a lawyer's analysis, not a search result.

The decision: send counsel the hard carbon patents with their independent claims, and drop the lithium-only results from the reading list.

3. Competitor patent watch: who is actually filing

The problem: a competitive intelligence analyst wants to know what QuantumScape patented in 2026 on solid-state separators.

What we ran: a 2026 search that named the company, then a search by company name with no date.

named = valyu.search(
    "QuantumScape solid-state battery separator patents",
    included_sources=["valyu/valyu-patents"],
    start_date="2026-01-01",
    max_num_results=10,
)
by_company = valyu.search(
    "QuantumScape Battery, Inc. patents",
    included_sources=["valyu/valyu-patents"],
    max_num_results=10,
)

What came back: the 2026 search returned 10 solid-state separator and electrolyte patents, none of them QuantumScape's. GM Global Technology Operations held 4, granted between January and September 2026, Murata 2, and Celgard, Samsung SDI and BMW one each. The search by name returned six QuantumScape patents granted between 2016 and 2024.

What we checked at the source: the assignee on each record.

The catch: a company name in the query steers relevance; it doesn't filter by owner. The first search found who is most active on the technology this year, which is a useful answer, but not the one asked. To watch one company, filter results on the assignee field, and confirm gaps with the USPTO's own assignee search before concluding that a competitor filed nothing.

The decision: add GM to the watch list for solid-state separators, and track QuantumScape with an assignee filter rather than a keyword.

4. Compound lookup: confirm identity before you order

The problem: a formulation chemist specifying LiFSI, lithium bis(fluorosulfonyl)imide, for an electrolyte needs its identifiers for a purchase order and safety review.

What we ran: a search of PubChem.

compound = valyu.search(
    "lithium bis(fluorosulfonyl)imide LiFSI",
    included_sources=["valyu/valyu-pubchem"],
    max_num_results=3,
)

What came back: one structured record: PubChem CID 86277430, formula F2LiNO4S2, molecular weight 187.1, CAS number 171611-11-3, with SMILES, InChIKey and synonyms.

What we checked at the source: PubChem's own API. The CID, formula and weight match, and a name search on PubChem returns the same single CID.

The catch: PubChem gives identity and computed properties, not the measured properties a formulator needs. Several computed fields were empty for this salt, and conductivity, purity grades and water content come from papers and supplier specifications.

The decision: put the CID and CAS number on the purchase order and safety review, and take conductivity and impurity limits from the supplier's specification.

5. State of the art: which record is the record

The problem: a strategy team writing a solar technology brief needs the current record efficiency for perovskite-silicon tandem cells.

What we ran: a web search, an arXiv search, and the Contents API on three of the articles it found, to get the certifier and conditions.

news = valyu.search(
    "record certified efficiency perovskite-silicon tandem solar cell 2026",
    included_sources=["web"],
    max_num_results=8,
)
details = valyu.contents(
    urls=[
        "https://www.pv-magazine.com/2026/07/15/longi-sets-new-world-record-with-35-5-efficient-perovskite-silicon-tandem-cell/",
        "https://www.pv-magazine.com/2026/11/06/longi-achieves-35-31-efficiency-in-perovskite-silicon-tandem-solar-cell-based-on-self-assembled-molecules/",
        "https://jsipe.com/blog/perovskite-silicon-tandem-solar-cell-hits-record-30-77-efficiency-2026-10-02",
    ],
    summary="State the efficiency, the cell area, who certified it, the date, and whether the article calls it a world record. Quote the sentence about certification.",
)

What came back: three 2026 figures:

The arXiv search returned research on tandem cells but no 2026 record figures.

What we checked at the source: the original articles, which give the certifier and conditions above.

The catch: "record" depends on the category: cell area, two-terminal or four-terminal, and who certified it. The 30.77% article doesn't name a certifier or say which category it leads, so it can't be compared with 35.5%. And preprint servers lag records announced at conferences, so arXiv alone misses them.

The decision: quote 35.5% as the record for a two-terminal perovskite-silicon tandem cell, with ESTI as certifier and July 2026 as the date, and check it against NREL's best research-cell efficiency chart before publishing.

6. Public funding: who the government is paying

The problem: a grants manager wants to know which companies have won federal awards for solid-state battery work, to find partners and gauge competition for funding.

What we ran: a search of federal awards on USAspending.

awards = valyu.search(
    "solid-state battery",
    included_sources=["valyu/valyu-usaspending"],
    max_num_results=10,
)

What came back: 10 award records with recipient, amount and agency. The largest was $2.09 million to Ion Storage Systems from the Department of Defense. Others included a $749,899 NASA contract to Solid Power and awards to Navitas Advanced Solutions Group and Carbon Solutions.

What we checked at the source: the Ion Storage Systems award on USAspending.gov. It was a Defense Contract Management Agency contract for production development of next-generation solid-state and lithium metal batteries, running from July 2020 to July 2022.

The catch: the records carry amounts but not dates or descriptions, and the largest award had ended more than four years earlier. A longer question, "Department of Energy solid-state battery research awards", returned nothing, so keep award searches short.

The decision: open each award on USAspending to see its dates before contacting partners, and search the names of active funding programmes for current opportunities.

What the six jobs cost

Patents use 1.28 credits per result and ChemRxiv 0.8, while arXiv uses 0.05 and web 0.1. Patent searches were most of the cost. At pay-as-you-go rates of 1,600 credits for $10, the 61 credits cost about $0.38.

Go further with DeepResearch: a technology brief in one task

The jobs above each answer one question. An R&D team's actual output is a technology brief or a patent landscape for a go or no-go decision. DeepResearch writes these. It plans the research across preprints, patents, compound data and the web, and returns a cited report with files.

task = valyu.deepresearch.create(
    query="Write a technology brief on air-stable sulfide solid electrolytes for all-solid-state batteries.",
    mode="standard",
    research_strategy="Use ChemRxiv and arXiv for 2025 and 2026 results, USPTO and EPO for patents with their assignees, and the web for company announcements. Report every performance figure with its test conditions. Filter patents by assignee field, not by name in the text.",
    report_format="A technology brief: State of the art, Leading approaches, Who is patenting, Open problems, Recommendation. Every figure cited with its conditions.",
    search={"included_sources": ["valyu/valyu-chemrxiv", "valyu/valyu-arxiv", "valyu/valyu-patents", "valyu/valyu-patents-epo", "web"]},
    deliverables=["An xlsx of reported results with material, ionic conductivity, air-exposure conditions, retention and source"],
)
 
result = valyu.deepresearch.wait(task.deepresearch_id)
print(result.output)  # cited report in markdown
for file in result.deliverables:
    print(file.type, file.url)  # signed download link

The research strategy carries the lessons from the six jobs: report every figure with its conditions, and filter patents by assignee.

Start from a ready-made workflow

Valyu's curated DeepResearch workflows include several for R&D and technical strategy, each with set inputs and file deliverables:

To run one from code, pass its slug as workflow_id:

task = valyu.deepresearch.create(
    workflow_id="legal-patent-landscape",
    workflow_params={"technology": "hard carbon anodes for sodium-ion batteries"},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

For worked examples, see the chemistry and materials workflows and running a workflow in the docs.

Where AI agents stop in R&D

The jobs above show what an agent can do from published sources. Four limits shape how to use one:

• Search isn't a legal opinion. An agent can find and rank patents; whether a product infringes depends on claims, jurisdictions and legal status, which counsel decides.
• Relevance isn't ownership. Company names in a query steer results but don't filter them. Filter on assignees, and confirm absences in the patent office's own search.
• Numbers need their conditions. Efficiencies, conductivities and capacity retention only compare when the test conditions match. Ask for the conditions with every figure.
• Your own data stays separate. Lab notebooks, internal reports and confidential results aren't in public sources. The agent covers the outside world; your team connects it to what you know.

How to set up an R&D research agent

If you don't want to write code as shown above, there is another option to connect Valyu to Claude or ChatGPT as a custom connector at https://mcp.valyu.ai/mcp and ask the questions above in plain English. To build your own agent, see ChemRxiv in Valyu, semantic patent search across the USPTO and how to integrate research papers into your AI agents.

Three habits made the jobs above reliable:

• Filter on structured fields. Assignee, date and source fields do what words in the query can't.
• Ask for conditions with every number. A figure without its test conditions or certifier isn't comparable.
• Check the record before you act. Open the patent, award or compound record behind any result that drives a decision.

FAQ

What is AI for R&D?

AI for R&D is the use of language models and AI agents to search and read the evidence around research and development: scientific preprints and papers, patents, compound databases, funding records and industry news. Most of the value today is in faster, cited technology landscapes, patent searches and state-of-the-art checks, not in replacing experiments.

What are AI agents for chemistry and materials research?

AI agents for chemistry and materials research are language models that search sources such as ChemRxiv, arXiv, USPTO and EPO patents and PubChem, read what they find and return cited answers. They handle the literature, patent and compound lookups that come before a scientist's judgement and the lab work.

What is the best AI for chemistry research?

AI chemistry research tools fall into two groups. Enterprise platforms such as CAS SciFinder and other chemistry-specific products search their own curated databases. Search APIs such as Valyu let you build your own agent, or connect Claude and ChatGPT to ChemRxiv, arXiv, PubChem, ChEMBL, USPTO and EPO patents and the web, with a source for every result.

Can AI do a patent search or freedom-to-operate analysis?

AI can run the first-pass patent search, not the legal analysis. In a November 2026 test, an AI agent returned 19 recent US and European patents on sodium-ion hard carbon anodes with full text and claims, but 3 of the 9 US results were lithium-ion patents. Counsel reviews the claims, jurisdictions and legal status before any freedom-to-operate opinion.

How do I track a competitor's patents with AI?

Filter by assignee rather than putting the company name in the query. In a November 2026 test, asking for QuantumScape's 2026 separator patents returned none of QuantumScape's, while GM held 4 of the 10 most relevant 2026 grants. Confirm any apparent gap with the USPTO's own assignee search.

Can AI find the current state of the art in a field?

Yes, if it searches the web and checks the conditions behind each figure. In a November 2026 test, web search found three 2026 perovskite-silicon tandem efficiency figures, from 30.77% to 35.5%, certified by different labs for different device classes. arXiv alone returned no 2026 record figures, because records are often announced before they appear as preprints.

What data sources does an AI R&D agent need?

An AI R&D agent needs preprints and papers (ChemRxiv, arXiv, PubMed), patents (USPTO, EPO), compound and bioactivity data (PubChem, ChEMBL), funding records (USAspending, NIH RePORTER) and the web for company announcements. Valyu searches all of these through one API.

How much does AI-assisted R&D research cost?

In a November 2026 test, six R&D jobs with Valyu used 12 API calls and 61 credits, about $0.38 at pay-as-you-go rates. Patent results cost 1.28 credits each and made up most of the total. A standard DeepResearch technology brief uses 100 credits, about $0.63.

Can AI replace R&D scientists?

No. AI agents speed up the reading around R&D: landscapes, patent searches, compound lookups and state-of-the-art checks. Designing experiments, interpreting results and deciding what to build remain with scientists, and legal conclusions on patents remain with counsel.

Start with one job

Pick the job your team repeats most often, such as the monthly preprint scan or the first patent search on a new idea, and run it on a topic you already know well. Compare the agent's results with your own.

Create a Valyu account for free credits every month, up to 2,000 with a work email. Then connect it to Claude or ChatGPT at https://mcp.valyu.ai/mcp, or browse the data sources. For research in drug discovery, see AI for Drug Discovery.