AI as a Lab Partner: How Machine Learning Is Changing Scientific Research

Scientists have used computers for simulation and statistics for decades. What has changed recently is that machine learning systems are contributing to the creative parts of research: proposing structures, ranking hypotheses, designing experiments and reading literature at a scale no human can match.

The clearest signal of this shift came when the 2024 Nobel Prize in Chemistry recognised work on computational protein design and protein structure prediction, including the AlphaFold system. A review in Nature on scientific discovery in the age of artificial intelligence surveys how AI now supports nearly every stage of the research cycle, from data collection to hypothesis generation.

My own background in Geomatics Engineering gave me an early taste of this. Spatial analysis has long relied on algorithms to extract patterns from satellite imagery and survey data that would be impossible to process by hand. Modern deep learning has simply extended that reach into far more domains.

It is useful to break the contributions down by stage:

• Structure prediction: AlphaFold predicts a protein's three-dimensional structure from its amino-acid sequence with accuracy that, for many proteins, approaches experimental methods. That turns a task that could take a lab months into a computation.
• Candidate generation and screening: in materials science, graph neural networks trained on crystal structures can predict stability across vast numbers of hypothetical compounds. DeepMind's GNoME work, described in Nature, reported discovering a large number of candidate stable crystals, a portion of which were subsequently synthesised by independent researchers.
• Surrogate models: expensive physics simulations, such as climate or fluid dynamics models, can be approximated by neural networks that run orders of magnitude faster, enabling broader exploration of scenarios.
• Active learning and autonomous labs: a model chooses the next experiment most likely to be informative, a robot runs it and the result updates the model. This closed loop can explore large design spaces efficiently.
• Literature synthesis: language models help researchers search, summarise and connect findings across thousands of papers, though they must be used carefully.
• Data analysis: automated image segmentation in microscopy, anomaly detection in particle physics and species identification in ecological surveys all reduce manual labelling burdens.

The pattern that works is generate, filter, verify. AI is excellent at generating candidates and filtering them by predicted properties. Verification still requires experiment or rigorous simulation.

1. Frame the question narrowly. "Find a better catalyst" is too vague. "Rank these candidate compositions by predicted activity for this reaction under these conditions" is tractable.
2. Curate training data carefully. Scientific datasets often contain systematic biases from which experiments were historically easy or fashionable. Negative results are frequently missing, which skews models towards optimism.
3. Quantify uncertainty. Ensembles, Bayesian methods or conformal prediction give scientists a sense of where the model is extrapolating. A prediction without an uncertainty estimate is hard to act on.
4. Validate experimentally. Hold out a set of predictions for blind testing, and report the hit rate honestly.
5. Close the loop. Feed experimental outcomes, including failures, back into the model.

Climate science offers a useful example. In my work with StrikeClimate, the value of models depends heavily on combining remote sensing, ground observations and domain expertise. Machine learning can downscale coarse climate projections or fill gaps in sensor networks, but its outputs only matter if they are validated against field measurements.

From a software perspective, the infrastructure around the model often determines success more than the model itself. Experiment tracking, versioned datasets, containerised environments and clear provenance for every prediction allow a lab to answer the simple question that peer reviewers will ask: if we rerun this, do we get the same answer? Treating research code with the same discipline as production code, including tests, code review and continuous integration, pays off quickly when results need to be defended or extended by the next cohort of students.

AI-assisted science brings real risks that researchers are actively debating.

• Reproducibility: models trained with undocumented data splits, random seeds or proprietary datasets can be difficult to reproduce. Sharing code, weights and data, where possible, matters more than ever.
• Data leakage: if test structures are closely related to training structures, reported accuracy can be inflated. Careful splitting, for example by sequence similarity or time, is essential.
• Hallucinated literature: language models can invent citations or misstate findings. Every claim surfaced by an LLM must be traced to the primary source.
• Prediction is not understanding: a model that predicts well may not explain why. Interpretability work and mechanistic follow-up remain part of science.
• Compute inequality: training frontier scientific models requires resources concentrated in a few countries and companies. Open releases, such as the public AlphaFold structure database, help, but they do not eliminate the gap.
• Dual use: tools that design useful molecules could, in principle, be misused. Responsible release practices and screening are part of the field's governance conversation.

African universities and research centres face constraints in funding, equipment and compute, but AI-for-science also creates openings.

• Use open models and databases. Freely available structure predictions, pretrained materials models and open satellite data lower the barrier to world-class work.
• Focus on local problems. Neglected tropical diseases, crop varieties suited to local soils, groundwater mapping and urban flooding in cities like Lagos are areas where local data and context give African researchers an advantage.
• Collaborate across disciplines. Pair domain scientists with software engineers and data scientists. In collaborative AI projects I helped coordinate through Omdena, mixed teams consistently produced better-framed problems than purely technical ones.
• Invest in research software engineering. Clean, tested, reproducible code is a research output in its own right, and good engineers are as important to a lab as good instruments.
• Share data responsibly. Locally collected datasets are valuable; publishing them with clear licences raises the visibility of the institutions that produced them.

• AI now contributes to hypothesis generation and experiment design, not just data analysis.
• Structure prediction, materials discovery and surrogate models are among the most mature applications.
• The reliable pattern is generate, filter, verify, with experiments remaining the final arbiter.
• Uncertainty estimates, careful data splits and reproducible code are non-negotiable.
• Never trust an AI-surfaced citation without reading the primary source.
• African institutions can lead by applying open models to local problems with strong interdisciplinary teams.

1. Highly accurate protein structure prediction with AlphaFold — Nature
2. Scaling deep learning for materials discovery — Nature
3. Scientific discovery in the age of artificial intelligence — Nature
4. The Nobel Prize in Chemistry 2024 — NobelPrize.org

Cover image: original illustration created for this article. Cited works remain the property of their respective authors and publishers.