Where AI fits in scientific discovery
AI can support a research workflow by proposing candidates, predicting properties, prioritizing experiments, or interpreting measurements. These roles have different evidence requirements. A plausible molecule, a predicted structure, a successful synthesis, and a useful product are separate results.
This guide uses published research examples to explain those distinctions and offers a practical framework for evaluating an AI-assisted research project. It does not claim that all laboratories can reproduce a paper's results or that an AI prediction proves safety, efficacy, material novelty, or commercial value.
Distinguish generation, prediction, and experimentation
A generative model proposes an object such as a sequence or structure. A predictive model estimates a property or outcome for an input. An experimental system measures what happens under specified conditions. A workflow may combine all three, but its claims should identify which stage produced the evidence.
Candidate generation must respect domain constraints and feasibility. Prediction requires an appropriate evaluation set and uncertainty assessment. Experimentation requires calibrated instruments, controlled procedures, relevant measurements, and interpretation by qualified researchers. Improving one stage does not automatically improve the complete process.
Example: biomolecular structure prediction
The 2024 AlphaFold 3 paper describes a diffusion-based model for predicting structures of complexes containing proteins, nucleic acids, small molecules, ions, and modified residues. Its evaluation concerns structural prediction across specified benchmarks. Read the methods and limitations alongside the reported comparisons.
A predicted interaction structure can inform a research hypothesis. It does not, by itself, establish binding under the intended experimental conditions, biological activity, toxicity, or a clinical benefit. Those questions require their own evidence. Avoid translating a structure-prediction result into a claim that a drug is effective or ready for use.
For a research integration, record model version, input preparation, candidate selection, confidence outputs, and downstream validation. Define what researchers may conclude from the result and which follow-up measurement is needed. A useful interface preserves this context rather than displaying a prediction as a confirmed observation.
Example: closed-loop materials experiments
The A-Lab study describes a platform combining computation, historical literature, machine learning, active learning, and robotics for solid-state synthesis of inorganic powders. It illustrates an experimental loop: propose a recipe, execute an experiment, analyze its outcome, and use the evidence to inform another attempt.
The January 2026 author correction matters when interpreting this example. The authors clarified that original novelty language concerned materials new to the prediction platform, not necessarily new to science. They also reported confirmation of 36 of 40 originally reported successes, with four inconclusive identifications, and removed a compound from the discussion because it had been included in training data.
This is a concrete reason to inspect corrections, identification methods, and training overlap before repeating a discovery claim. Cite the corrected record and explain the scope of the evidence. A demonstration of a particular platform is not proof that any automated laboratory operates without ambiguity or oversight.
Choose a research question and a baseline
Start with a bounded question: which candidates should be measured next, which property should be estimated, or which experimental condition should be compared? Define the property, measurement procedure, constraints, and acceptable uncertainty. The question should be precise enough to decide whether the workflow helped.
Compare with a credible baseline such as the existing expert selection process, a simple statistical model, or a conventional search strategy. Use the same candidate pool, resource limits, and success definition. A comparison against an artificially weak alternative does not show that the system improves the actual research process.
Separate model performance from workflow performance. A prediction error metric is useful for a predictor; validated candidates per experimental budget may matter more to the laboratory. Record both when they are relevant and explain how they relate to the research decision.
Preserve data provenance and experimental context
Keep identifiers, units, measurement conditions, instrument settings, preparation methods, and uncertainty with each observation. Record whether a value was measured, simulated, inferred, or extracted from literature. Mixing these categories without context can produce misleading training targets.
Document exclusions and missing values. Failed experiments can be informative, but failure labels need a clear meaning: synthesis failure, measurement failure, unsuitable purity, or an operational interruption are different events. Do not train the model to treat an equipment fault as evidence that a candidate is intrinsically unsuitable.
Track data licenses, authorized access, and permitted uses. Preserve the source and transformation history so a researcher can investigate a surprising prediction or recreate a dataset. Model output should not erase the provenance of the evidence used to produce it.
Test beyond closely related training examples
Choose evaluation splits that represent the intended use. Randomly separating nearly identical compounds, repeated measurements, or related sequences can overstate performance on genuinely unfamiliar candidates. Consider grouping by the relevant family, experimental campaign, or time period when that matches the deployment question.
Keep final evaluation data separate from model selection and prompt development. Check duplicates and known training overlap where information is available. If overlap cannot be established, disclose that limitation rather than describing the evaluation as definitively independent.
Evaluate the regions in which the system will be used, including sparse or unusual inputs. Average error alone may conceal a failure for an important class of candidates. Record uncertainty behavior and define when the model should abstain or require additional measurement.
Keep physical execution within reviewed boundaries
An experiment proposed by a model should pass the laboratory's authorized review and execution process. Validate instrument capability, permitted conditions, materials compatibility, and required approvals before scheduling work. The responsible laboratory team must define these boundaries.
Test interruptions, sensor failures, conflicting measurements, and incomplete records. Establish who can stop a run and how partially completed experiments are reconciled. A closed-loop design needs explicit stopping conditions and escalation paths; repeated iteration is not a guarantee of scientific progress.
Separate the model's proposal from the equipment controller and its protections. Log proposed actions, validation decisions, executed settings, measurements, and human interventions. These records support both incident investigation and scientific interpretation.
A hypothetical pilot for candidate prioritization
Suppose a materials team already measures a defined property for a constrained family of candidates. It wants to compare an AI ranking with its existing selection process. This is an illustrative project design, not a reported TensorBlue experiment or a claim about achievable performance.
The team freezes a candidate pool, defines measurement conditions and a success threshold, and assigns comparable experimental budgets. It records the selection method before observing outcomes. Researchers review measurement quality and repeat important results using the agreed procedure.
The pilot reports verified outcomes, unsuccessful attempts, instrument interruptions, time, and costs. It also identifies which candidate families were not tested. Expansion depends on that evidence and operational readiness, rather than on the appearance of promising model scores.
Make results reproducible and claims reviewable
Retain code and model versions, dataset snapshots, preprocessing, seeds where applicable, configuration, and experimental protocols. Document dependencies and the environment needed to reproduce the computational steps. Preserve measurement artifacts and the analysis that supports a conclusion.
Write conclusions at the level the evidence supports. Distinguish a predicted candidate from an experimentally confirmed result and a reproduced result from a production-ready process. State unresolved uncertainty, corrections, and the conditions under which the result was obtained.
Explore our AI consulting services for defining an evaluation plan, read the MLOps guide for versioning and monitoring, and see AI for manufacturing for operational integration considerations. Contact TensorBlue with the research question, available data, measurement process, and decision the proposed system should support.