How AI is starting to change scientific modeling
In a guest essay on Anthropic's Science Blog, Tübingen researchers Jakob Macke, Daniel Gedon and Stefan Wahl split scientific modeling into three jobs and argue AI should return a range of plausible models, not one best fit. A summary, plus how to try simulation-based inference.
Updated November 20, 2026: Expanded with the essay's three steps in detail, its caveats, and how to try simulation-based inference yourself.
Anthropic's Science Blog published Teaching AI to build scientific models on November 6, 2026. It is a guest essay by Jakob Macke, Daniel Gedon and Stefan Wahl of the ELLIS Institute Tübingen and the University of Tübingen, all grantees of Anthropic's AI for Science program who build AI tools for scientific modeling. It is not a product announcement but an argument about what AI should hand back to scientists, and it reads well as a design brief for anyone building research agents.
The example behind the argument
The authors start with Hodgkin and Huxley's 1951 model of how a squid nerve cell fires. Huxley solved its equations on a hand-cranked calculator, which took more than three weeks. To fit the data, the pair had to raise one variable to the fourth power without knowing why. They cautiously guessed that a membrane gate might need several charged particles to move before it opened. The model predicted impulse speed relatively accurately, the pair won the 1963 Nobel Prize, and decades later the structure of a related potassium channel turned out to have four identical pieces around a central pore. The guess carried weight only because simpler versions of the model had failed. An AI could plausibly find such an equation in an afternoon today, the authors write, but insight comes from knowing which alternatives don't work.
Three jobs in modeling
- Building: deciding what the model is made of and what shape its equations take.
- Constraining: tuning the model's adjustable parameters to fit real measurements.
- Interpreting: working out what the model says about the real system.
Scientists loop through these steps repeatedly, and AI is now involved in all three. The essay says machines have arguably made the most progress on constraining, and that interpretation is the hardest step to hand over.
Constraining: many answers fit the same data
The authors compare fitting parameters to a $120 restaurant bill with no itemised receipt: many combinations of orders add up to the same total. Likewise, many combinations of gate parameters can produce nearly identical electrical activity in a nerve cell. Bayesian inference gives the whole range of parameter sets that fit, but it needs a formula for how likely the measurements are, which is often impractical. Simulation-based inference (SBI) needs only a simulator, parameters in and data out. You run it many times with random parameters, perhaps 100,000 times for a model like Hodgkin-Huxley and millions for larger ones, then train a neural network to map an output back to the parameters that could have produced it. The simulations can take weeks, but afterwards inference on new data takes seconds. The essay's example is working out roughly where a neutron-star collision happened, fast enough to point telescopes at it; it also cites uses in cosmology, structural biology, geophysics and nerve-cell models.
Building: search the space of models
The same problem appears one level up. Several AI systems from the past two years use a language model to write candidate model code, fit it, compare it with data and revise it. But many differently shaped models can explain the same data, and two runs of an agent may return two different, equally good answers. The authors propose treating model building like parameter fitting: propose many candidates, weight each by how well the data supports it (with less credit for those needing a lot of flexibility), and keep refining the strong ones. If every surviving model shares a feature, the data likely favours it; where survivors disagree, that points to the next experiment.
They add a caveat. So far, these systems have mostly been tested on rediscovering models scientists already know. Evidence that they work on problems with no known answer is sparse.
Interpreting, and models too big to read
Interpretation is hard to automate because there is no clear objective for understanding. The essay mentions two proposals: judge an AI's understanding by how well it predicts a system's response when disturbed, or by whether its output helps human scientists understand the problem. The authors also look ahead to models too large to write by hand, such as work built on the complete wiring diagram of the fruit fly brain, shown in the essay's figure with 124,635 cells. At that scale there is no single correct model, so the useful question becomes what all data-compatible models agree on.
Try simulation-based inference
The essay doesn't point to a toolkit; it links SBI to a research paper. Our own suggestion for trying it is sbi, an open-source Python package for simulation-based inference, several of whose methods come from papers co-authored by Macke. It needs Python 3.11 or later (a GPU isn't necessary), and its README recommends installing with uv, then running a short test:
uv pip install sbi
# test the install (from the README), in Python:
from sbi.examples.minimal import simple
posterior = simple()
print(posterior)For your own model, the README's minimal pattern is to define a prior, simulate parameters and outputs, pass them to NPE (neural posterior estimation), train it and build a posterior. It offers amortised and sequential methods plus validation and plotting tools. The README also points coding agents to an llms.txt index of its documentation, useful if you ask Claude to wire up your own simulator, and recommends a 2025 tutorial paper, Simulation-based inference: A practical guide.
Our take
In our reading, the essay's most practical point for anyone building research tools on Claude is output format. The authors argue that a single best fit hides the other answers that would have fit just as well. For tools, that suggests returning a ranked family of candidate models, so the scientist sees what the data actually pins down. That is harder to build and present, but in our view it is what would make an AI collaborator trustworthy in science rather than merely fast.
Source: Teaching AI to build scientific models (Anthropic), sbi: Simulation-Based Inference (GitHub README)
This story first appeared in the claudefox daily briefing for November 7, 2026.