Lecture 36: Science — How AI Supports Scientific Research

Part 1 (20 minutes) — Overview, analogies and examples

In modern research, Artificial Intelligence (AI) supports scientists in three fundamental ways: (1) by automating repetitive or laborious subtasks, (2) by analyzing very large datasets, and (3) by proposing new candidates that can be tested experimentally. A clear analogy is that of a microscope combined with a librarian and an assistant: the microscope expands our perception (measurement data), the librarian filters and organizes existing knowledge (literature and data analysis), and the assistant suggests promising next steps (e.g., which compound to synthesize or which material to test).

Concrete, well-documented examples illustrate these three roles:

Protein structure prediction: DeepMind published results in 2021 on AlphaFold, an AI system that can predict structures for many proteins with high accuracy. These predictions reduce the need for lengthy experimental methods for an initial structural hypothesis and thereby accelerate biomedical research steps, for example in understanding functional mechanisms or in drug design (see Jumper et al., 2021).

Automated laboratory research: There are approaches in which robotics, measurement apparatus and AI interact in closed loops: the AI selects experimental conditions, robots perform the experiments, measurement results flow back and the AI adjusts its recommendations. Such so-called "self-driving laboratories" have shown that the search for optimal materials or process conditions can be significantly accelerated (see MacLeod et al., 2020).

Computer-assisted synthesis planning: In chemistry, AI systems support retrosynthetic planning, i.e., the question how a desired molecule can be made from available starting materials. Models that have learned large numbers of published synthesis routes can provide sensible suggestions and thus reduce planning time for chemists (see Segler et al., 2018).

Text and literature analysis: During epidemics and for fast-moving research topics, aggregating large literature databases and machine text-mining has helped researchers identify relevant results more quickly and formulate hypotheses. One example is the provision of large, centralized literature datasets that enable automated analyses (see CORD-19).

These examples show: AI is not a panacea, but in many areas an accelerator — especially where large datasets, recurring patterns, or clear optimization goals are present.

Part 2 (20 minutes) — Technical terms, methods and scientific context

In this section I introduce central technical terms and explain their significance for the scientific use of AI. I refer to literature that describes practical examples and methodological foundations.

Surrogate models and predictive models

Surrogate models are statistical or ML models that approximate an expensive experiment. Instead of repeatedly using an expensive measurement setup, a surrogate can first roughly predict which parameters are promising; the most promising candidates are then tested experimentally. This principle underlies many "self-driving lab" approaches (MacLeod et al., 2020) and also material or drug searches where experiments are costly.

Inverse design

Inverse design refers to the search for designs (e.g., molecules, materials) that satisfy certain desired properties. Machine learning methods — particularly generative models — can propose a large number of potential candidates. Gómez-Bombarelli and colleagues described approaches to map molecules into continuous representations and thereby generate new candidates that are then evaluated and synthesized (Gómez-Bombarelli et al., 2018).

Active learning and closed optimization loops

Active learning here means that the system deliberately selects the most informative next experiments in order to maximize knowledge gain. In practice, one often combines Bayesian optimization or similar strategies with robotic experiments: the AI suggests experimental parameters that would reveal particularly much about the underlying system. Such closed loops are central to accelerated discovery (MacLeod et al., 2020).

Retrosynthesis and symbolic-statistical hybrids

For planning chemical syntheses, a combination of data-driven models (e.g., neural networks) and symbolic rules (chemistry-specific reaction catalogs) is often used. Segler et al. (2018) showed how neural models together with symbolic AI can generate practicable synthesis plans.

Interpretability, causality and the limits of statistical learning

Important for science is not only predictive accuracy but also understanding why a model makes a prediction. Interpretability is an active research field; without it, it often remains unclear whether a model has learned real causal relationships or only correlations. Books and works on causality (e.g., Peters et al., 2017) emphasize that causal conclusions from purely observational data are only possible under assumptions. This means for science: models can suggest hypotheses, but causal tests must be carried out experimentally.

Bias, data quality and reproducibility

Machine learning models are only as good as the data on which they were trained. Issues such as lack of representativeness, biased training data and undocumented preprocessing can lead to misdirection. Critical contributions like "On the Dangers of Stochastic Parrots" (Bender et al., 2021) highlight the risks of large, undifferentiated datasets. In research, therefore, data provenance, quality and processing must be transparently documented and verifiable to ensure reproducibility.

What AI does well so far — and what not

From the works cited, a scientific consensus emerges in the literature: AI can quickly and reliably find patterns in large datasets, generate hypotheses and automate routine tasks; however, it currently cannot "prove" complex, novel scientific theories or fully autonomously generate them without human planning, critical review and experimental validation. Approaches like the "Robot Scientist" demonstrate automation and hypothesis generation, but final confirmation remains experimental and expert-reviewed (King et al., 2009).

I explicitly point out where uncertainties remain: in the long term it is open to what extent AI can truly discover novel, fundamental theories. This depends both on methodological advances (e.g., better causal learning, explainable models) and on the availability of high-quality, well-documented data. This question is the subject of active research, not a settled consensus.

Part 3 (10 minutes) — Concrete applications, limits and thought exercises

Concrete applications (brief overview)

Protein structure prediction: AlphaFold (Jumper et al., 2021) provides high-quality structure predictions for many proteins, accelerating understanding of protein function and potential drug approaches.

Materials and process optimization: Self-driving laboratories use AI-driven loops to optimize parameters for, e.g., thin films or catalysts; this can reduce the number of experiments and development time (MacLeod et al., 2020).

Computer-assisted synthesis planning: AI systems support chemists in planning practical synthesis routes and can thus shorten development cycles (Segler et al., 2018).

Literature and data analysis: Centralized data collections and text-mining have helped in fast-moving research situations to aggregate relevant evidence quickly (CORD-19).

Key limitations — stated precisely

- Causality: Models mostly find correlations; causal conclusions require additional experimental tests and/or causal model assumptions (Peters et al., 2017).
- Data dependence: Biased, incomplete or mislabelled data lead to faulty results; careful data auditing is necessary (Bender et al., 2021).
- Interpretability: Many powerful models are hard to interpret; without interpretability scientific insight is more limited (Doshi-Velez & Kim, 2017 — see literature note for the state of the discussion).
- Reproducibility: Lack of disclosure of training data, hyperparameters or pipelines makes reproducing results difficult.
- Experimental validation: AI suggestions often require elaborate experimental testing; the bottleneck sometimes shifts from "finding ideas" to "confirming and understanding".

Small thought exercises for teaching / self-assessment

1) You are given a model that predicts a certain substance increases the activity of an enzyme. Describe a minimal experimental plan to test whether this is a causal effect or a correlation. (Hint: Which controls, doses and replicates do you need?)

2) Assume an ML model suggests a new catalyst combination. What information about the training data would you request before investing resources into synthesis? (Think of data sources, measurement protocols, limits of the parameter range.)

3) A fully trained model achieves 95% prediction accuracy on the test dataset. Formulate three hypotheses why this figure could perform significantly worse in field application (e.g., domain shift). How would you test these hypotheses experimentally?

These tasks are intended to encourage not accepting the method as a black box, but to develop critical tests and validation plans — precisely what distinguishes good scientific practice when using AI.

Final pragmatic recommendation

For scientists today: use AI as an accelerator and exploration tool, but integrate robust validation and documentation steps into your workflow. Open data, clear documentation of procedures and close collaboration between domain experts and data scientists increase the reliability and scientific value of AI-supported research.