How AI Is Changing Scientific Modeling: From Manual Coding to Automated Exploration

Scientific models are a core tool for researchers to explain and predict real-world systems—from how a single neuron fires to what happens when neutron stars collide. Recent advances have changed the way scientists fit complex models to data. But a key bottleneck remains: scientists still typically have to write models by hand, and often must weigh multiple competing versions, each of which may explain the data in a different way.

The Current State and Bottlenecks of Scientific Modeling

According to the Science Blog, despite continued advances in data-fitting techniques, the process of building scientific models remains highly manual. Researchers must write equations or code describing system behavior by hand based on their domain knowledge, then test different hypotheses one by one. The limitation of this approach is that human imagination and time are both finite—you can only try model structures you can think of.

Even trickier is the problem of model selection. Given the same set of observational data, multiple models may exist that all provide reasonable explanations. Scientists need to judge which explanation is better supported by the data, which usually involves repeated fitting, comparison, and revision. This process is both time-consuming and susceptible to prior assumptions.

Model Selection and the Risk of Overfitting: When multiple models can reasonably explain the same dataset, scientists face the core problem of model selection. An overly complex model can perfectly fit existing data yet predict new data poorly—this is overfitting. Statisticians have developed complexity-penalizing metrics such as the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC), as well as methods like cross-validation. But these methods only apply to comparisons within a given set of candidate models; if the candidates themselves have not been exhaustively enumerated, selection remains limited by the researcher's imagination. This is precisely the bottleneck that AI-assisted model exploration attempts to break through: it not only compares existing models but can also generate new candidate structures, thereby expanding the search space.

The Four Stages Where AI Enters the Modeling Workflow

In a guest article, Jakob Macke, Daniel Gedon, and Stefan Wahl describe how AI is beginning to play a role in these stages. Their core argument is that AI can write, run, fit, and revise scientific models as code.

The significance of this workflow lies in automating the loop of model exploration. AI does not merely fit a given model; it can explore many possible model structures and identify which explanations the data support. In other words, AI participates in a search over model space, not just parameter optimization under a fixed structure.

This capability corresponds to the two most labor-intensive parts of scientific modeling: first, proposing candidate model structures; second, judging which structure is more reasonable. When AI can generate and test large numbers of candidate models, researchers can focus their energy on problem definition and interpretation of results.

Symbolic Regression and Program Synthesis: Letting AI "write scientific models as code" corresponds technically to two directions: symbolic regression and program synthesis. Symbolic regression does not presuppose a functional form; instead, it searches for mathematical expressions that explain observations from data, with representative methods including genetic programming and sparse regression (such as SINDy). Program synthesis goes further, enabling systems to search for executable code fragments rather than just mathematical formulas. These methods allow AI to automatically propose candidate structures in model space, rather than merely fitting parameters of a fixed form. The challenges lie in the enormous search space, sensitivity to noise, and the fact that generated models are not necessarily scientifically interpretable—so domain knowledge is still needed for screening and validation.

From "Fitting Data" to "Exploring Explanations"

Over the past few years, machine learning applications in science have mainly focused on the fitting stage—using neural networks or other models to approximate patterns in data. But the path described by Macke et al. goes further: letting AI participate in the construction and revision of models themselves.

This brings a subtle but important shift. Traditionally, scientists first have a hypothesis, then write a model, then test it with data. AI-assisted modeling allows multiple hypotheses to be generated and evaluated in parallel, with data playing the role of a filter. Researchers no longer need to exhaustively enumerate all possible model forms in advance; instead, they can let the system explore structures that humans might overlook.

Of course, this does not mean scientists can completely let go. The article explicitly states that scientists still typically write models by hand, and AI's involvement is auxiliary. How to define the problem and how to judge whether a model is scientifically reasonable still depend on domain expertise.

The Positioning of AI for Science: AI for Science broadly refers to the research direction of applying machine learning to scientific discovery, covering protein structure prediction, materials screening, climate simulation, and more. Most current work focuses on using AI to replace or accelerate computation—that is, at the "fitting data" level. Letting AI participate in hypothesis generation and model construction touches the core of the scientific method—proposing testable explanations. The difficulty of this shift lies not only in algorithms but also in the philosophy of science: how should the interpretability and falsifiability of an algorithm-generated model be evaluated? Researchers need to establish new norms for judging whether AI-proposed models have scientific value, not just statistical goodness of fit.

Impact on Research Workflows

If AI can reliably write, run, fit, and revise model code, research workflows will change. The iteration speed of model building may accelerate, because the generation and initial screening of candidate structures can be partially automated. The way researchers collaborate may also change—model code becomes an object that AI can operate on and search, rather than merely a medium of communication among humans.

However, this path also raises new requirements. Exploring models as code means that code correctness, reproducibility, and interpretability become more critical. If an AI-generated model structure yields a conclusion supported by data, researchers need to be able to understand why it holds, not merely accept the fitting result.

From the information revealed in the blog, this work is still at an early stage, with the focus on demonstrating the possibilities of AI in the modeling workflow rather than providing a mature automated system. For researchers following the AI for Science direction, this offers a perspective worth tracking: AI's role in science is extending from a data-processing tool to a participant in the modeling process.

One-line Summary

AI is expanding from fitting data to automatically writing, running, and revising scientific model code, exploring model space.

Paragraph Summary

The article discusses a long-standing bottleneck in scientific modeling: model construction remains highly manual, with researchers needing to write equations or code by hand and weigh multiple competing explanations. Jakob Macke, Daniel Gedon, and Stefan Wahl propose that AI can write, run, fit, and revise scientific models as code, thereby automating the model exploration loop. This extends AI's role from parameter fitting under fixed structures to structural search in model space, helping researchers discover candidate models that might otherwise be overlooked. But scientists still remain responsible for problem definition and judgments of scientific reasonableness. This direction is still at an early stage, with the focus on demonstrating possibilities, while also bringing new requirements for code correctness, reproducibility, and interpretability.