AI-for-Science Catchup: Developments We've Been Tracking
Generative crystal design moves from research papers to usable tools, Elicit ships Research Agents and a public API, and autonomous lab platforms cross from industry into academic reach.
A roundup of what’s actually new and usable (or worth watching) across the fields this site covers — sourced from lab announcements, peer-reviewed papers, and tool changelogs, not general AI hype newsletters.
Headline: Generative crystal design has real tools now
Computational crystal structure prediction has been an active research area for years, but usable generative tools — ones outside experts’ private codebases — have been slow to arrive. That changed in 2025.
CrystalFlow (published in Nature Communications, 2025) uses a flow-based generative approach to propose novel crystal structures conditioned on composition. It represents the clearest peer-reviewed step toward generative crystal design that’s citable and accessible to researchers outside the original group.
DiffCrysGen (arXiv, 2025) takes a score-based diffusion approach and reports 2–3 orders of magnitude faster sampling than prior diffusion methods for crystal generation, which matters for researchers who want to run large screening campaigns without a cluster.
A third approach, PCCD, emphasizes synthesizability constraints during generation — addressing the longstanding problem that generative models for materials often propose structures that look good computationally but can’t actually be made in the lab.
Why this matters: These tools make crystal structure generation a realistic option for materials science researchers who previously needed either expensive DFT campaigns or connections to a group working on this problem directly. The peer-reviewed CrystalFlow result in particular gives you a citable Tier 1 source for benchmark comparisons.
Caveats: All three are research code, not polished tools — expect to work with Python repositories rather than a sign-up flow. The DiffCrysGen preprint had not yet appeared in a peer-reviewed journal as of this writing. Verify independently before treating arXiv benchmark numbers as settled.
Sources: CrystalFlow, Nature Communications; DiffCrysGen, arXiv
Also worth knowing
Elicit shipped several significant features since its last major update. The literature search tool now covers PubMed and ClinicalTrials.gov (added October 2025) in addition to its general academic corpus. In December 2025, Elicit introduced “Research Agents” — autonomous multi-step investigations that can run a structured literature search end-to-end with less step-by-step instruction from the researcher. A public API launched in March 2026, enabling programmatic access for research teams that want to integrate Elicit into larger workflows. The corpus now stands at 138 million papers. We’ve updated the Elicit tool page to reflect the current pricing and features.
Autonomous lab platforms have moved from prototype to scaled deployment. Three developments worth tracking together as a trend:
- Lila Sciences (Flagship Pioneering, unveiled March 2025, $550M funding) is the highest-profile entry — a fully autonomous biology lab built around robotic platforms and AI-driven experimental planning. Not independently accessible; included here because its scale signals where industrial investment is flowing.
- BayBE (Merck KGaA and University of Toronto Acceleration Consortium, open-source) is the most relevant to academic researchers right now — a Python library for Bayesian optimization over experimental parameter spaces, with built-in support for categorical chemistry parameters and molecular encodings. It’s the kind of lab-automation infrastructure a single research group can actually deploy. See our new BayBE tool page.
- Distiller (Berkeley Lab, released April 2025) streams electron microscope data in real time to the Perlmutter supercomputer for AI-driven analysis — a model for how national labs are building closed-loop experimental infrastructure around AI rather than building AI tools that researchers bolt onto existing workflows.
Sources: Elicit changelog (elicit.com); R&D World; Royal Society Open Science review of self-driving labs
Worth watching (research-stage / not yet independently usable)
IBM RXN for Chemistry substantially improved after integrating Thieme’s Science of Synthesis dataset. IBM Research reports approximately 9× improvement in retrosynthesis prediction accuracy and ~3× improvement in forward reaction prediction following this integration. Science of Synthesis is a curated expert-authored synthesis database that covers reaction classes underrepresented in the patent literature — the usual training source for retrosynthesis models. If retrosynthesis is part of your workflow, this warrants a re-evaluation if you last tested IBM RXN before this update. Source: IBM Research blog
Correction / status update
Elicit pricing and features: Our tool page has been updated to reflect the current state. The $12/month Plus tier is no longer available; the current Pro tier is ~$49/month. Research Agents, PubMed search, and the public API are all additions since our original page was written.
Have a development we should cover next month? Use the Submit page to suggest a story or share how you’re using one of these tools in your own research.