Kevin Maik Jablonka
We think AI has real scientific impact when it helps us develop new concepts and work with the data that actually matter to scientists.
I lead LAMALab at Friedrich Schiller University Jena. I want AI to do more than make predictions. Simply making more predictions can distract us from the scientific questions that matter. I believe its most important scientific use will be to deepen our understanding, help us develop new concepts, and give us recommendations we can act on. Building that path requires end-to-end learning and models whose conclusions we can understand and test. Our work has made oxidation states computable, revealed unexpected behaviour in an operating carbon-capture plant, and inferred molecular structures directly from raw spectra.
The Science of the Unaskable
Some questions remain out of reach because the concept, representation, or measurement they require does not yet exist. The task is to build that missing piece and carry it through to an experiment or decision. Build it on the measurement and the action rather than on an idealised structure, and the answer comes back actionable: that is how the unaskable turns into actionable understanding. The research programme sets out the path and the questions still open along it.
I want to know what a model has learned, which representation makes a conclusion possible, and where that conclusion fails. This connects to older questions about emergence and levels of description and about what we mean by molecular structure and shape.
Making this work requires chemistry, machine learning, psychometrics, and econometrics; I also work and teach with psychologists and social scientists. LAMALab created ChemPile, ChemBench, and MaCBench; through NFDI FAIRmat, we help make materials data machine-actionable. These resources let others inspect, test, and extend the work.
Science and people
I am in a privileged position, so I keep asking myself: How can I make the most significant contribution?
My two most important outputs are science and people; in fundamental research, people may matter even more. Science should change what we understand. People carry that understanding—and the tacit knowledge behind it—forward.
At LAMALab, scientific freedom rests on trust, kindness, and candour. We share unfinished ideas, data, code, and failures early. My role is to protect time for thought, ask hard questions, and help each researcher develop an independent direction. I want the lab to leave better questions, resources others build on, and people ready to lead.
Bio lengths
Kevin Maik Jablonka leads LAMALab at Friedrich Schiller University Jena. He develops AI as a scientific instrument: making chemical concepts measurable and turning raw observations into the next experiment. His work has made oxidation states computable, inferred molecular structures from spectra, and revealed unexpected behaviour in an operating carbon-capture plant.
Kevin Maik Jablonka leads LAMALab at Friedrich Schiller University Jena. He develops AI as a scientific instrument: making chemical concepts measurable and turning raw observations into the next experiment. His work has made oxidation states computable, inferred molecular structures from spectra, and revealed unexpected behaviour in an operating carbon-capture plant. He combines chemistry and machine learning with psychometrics and econometrics to make scientific measurement, uncertainty, and judgement explicit. LAMALab works with psychologists and social scientists to test where model and expert judgements agree, where they fail, and which evidence changes them. These tests connect model conclusions to experiments and decisions.
Kevin Maik Jablonka leads LAMALab at Friedrich Schiller University Jena. He develops AI as a scientific instrument: making chemical concepts measurable and turning raw observations into the next experiment. His work has made oxidation states computable, inferred molecular structures from spectra, and revealed unexpected behaviour in an operating carbon-capture plant. His programme spans end-to-end learning: from the measurements scientists actually make, through representations and models, to recommendations they can test. Actionable recommendations depend on measurements, representations, experiments, and the tacit knowledge connecting them; each link must remain open to scrutiny. Scientists should be able to determine what a model learned, challenge its representation, and locate its limits before acting on it. LAMALab created ChemPile, ChemBench, and MaCBench as shared infrastructure others can inspect and extend. Psychometrics, econometrics, psychology, and social science help the group study measurement and judgement in models and people. He is a Google Research Scholar. At LAMALab, scientific freedom rests on radical openness, trust, and kindness. Ideas, data, code, and failures are shared early; close collaboration passes on tacit knowledge while each researcher develops an independent direction. He wants LAMALab's legacy to be new understanding, resources others build on, and people who lead programmes of their own.
Recent work has appeared in Nature Chemistry, Nature Computational Science, Nature Communications, Science Advances, and the NeurIPS main proceedings; the full list says what each established. ORCID. 5,539 Google Scholar citations, h-index 27 (checked ). More than €4M in awarded funding as principal investigator or project lead.
Curriculum vitae
- —presentIndependent Research Group LeaderCarl Zeiss Foundation Research Group, Friedrich Schiller University Jena. Member of the Center for Energy and Environmental Chemistry Jena and the Jena Center for Soft Matter; affiliated with HIPOLE Jena, the Michael Stifel Center, and the Acceleration Consortium in Toronto. Co-Speaker of NFDI FAIRmat. Eleven doctoral researchers and one postdoctoral researcher; 26 appointments supervised since 2021, whose alumni have gone on to graduate study in the United States and to research roles in industry.
- —presentIndependent contractor, AI evaluation and red teamingOpenAI.
- —PhD, Chemistry and Chemical EngineeringEPFL; Dimitris N. Chorafas Foundation Award and top-8% thesis distinction. Work from that period.
- —MSc, ChemistryEPFL; 5.95/6.00, second-highest GPA across all EPFL master’s graduates and highest in chemistry.
- —BSc, ChemistryTechnical University of Munich; high distinction.
More than €4M awarded as principal investigator or project lead: a Carl Zeiss Foundation research group (2023–2029), an Open Philanthropy project on agent benchmarks for spectroscopy, the Intel–Merck “MatAssist” project, and a Google Research Scholar Award.
Within consortia: local scientific lead for the Helmholtz Foundation Model Initiative project SOL-AI, principal investigator on two projects in the Carl Zeiss Foundation’s Nano@Liver programme, principal investigator and work-package leader in CARE-O-SENE and its awarded follow-up, and a project in the DFG Research Training Group COIN. Competitive compute from the Gauss Centre for Supercomputing and an NVIDIA Academic Grant.
Headline speaker at the Faraday Discussion on Emerging Materials for Optoelectronics in Edinburgh; the Nature Conference on AI for Sustainable Materials in Xi’an; the Frontiers of Solid State Theory symposium at the Max Planck Institute for Solid State Research; the Lennard-Jones Centre in Cambridge; the Machine Learning and Chemistry seminar at Caltech; CECAM’s AIChemist flagship workshop in Lausanne.
Keynotes at Machine Learning Physique Chimie in Lyon and at the Federation of European Zeolite Associations; a plenary at the first LLM for Materials hackathon at Imperial College London; and a panel with the President of the Helmholtz Association at Helmholtz Tech & Industry Day.
Critical Aspects of AI at Friedrich Schiller University Jena, taught jointly with sociology: chemists and social scientists read each other’s literature and argue about what these systems are doing to research practice. Alongside it, a graduate seminar on current topics in cheminformatics.
Previously: machine learning at the MolSim winter school in Amsterdam, and research data management, computational carpentry and the electronic lab notebook at EPFL.
Open material: matextract.pub, gpmbook.lamalab.org, interactive spectroscopy tools, and the machine learning course. More on teaching and supervision.
- Co-Speaker of NFDI FAIRmat; co-leader of WG3 in the DAEMON COST Action.
- Area Chair for NeurIPS 2026; area chair for the AI4Mat workshops at NeurIPS 2024 and ICLR 2025.
- Co-organiser of the 2027 Faraday Discussion on Digital Chemical Discovery; guest editor for a special issue on LLMs in Digital Discovery.
- Initiator of the first global chemical-sciences LLM hackathon; organiser of CECAM’s L2M3 and MADICES workshops.
- Grant reviewer for the ERC, DFG, A*STAR, FWF, ScaDS.AI and NWO; journal reviewer for the Nature and Science families, PNAS and JACS.
- Represents HZB/HIPOLE in Helmholtz programme-oriented funding evaluations. Helmholtz Leadership Academy.
- — Elected member, Junge Akademie Mainz
- — Member, Foresight Institute Berlin; ICML Gold Reviewer
- — Google Research Scholar Award
- — Helmholtz AI Award for Best Digital Resource, for ChemPile
- — Best Research Environment, Die Junge Akademie and the Volkswagen Foundation
- — Finalist, Rising Star in Computational Materials Science
- — Dimitris N. Chorafas Foundation Award
- — CAS Future Leader
- — Alfred Werner scholarship, Swiss Chemical Society
- — Studienstiftung des deutschen Volkes
Before any of that, I feel privileged to have taken part in student science fairs — a silver medal at the International Conference of Young Scientists in 2013, third nationally at Jugend forscht in 2014 — under the guidance of the mentors at a student research centre.
Photos and source material
Headshots by Marina Romanova / HIPOLE Jena.
