When Cambridge University’s computer stopped working for six months in 1951, physiologist Andrew Huxley had to abandon plans to use it to test equations he and Alan Hodgkin had devised to describe electrical signals in squid nerves. Instead, he worked through the calculations on a hand-operated mechanical calculator, a task that took more than three weeks.
The equations sought to explain how nerve cells generate action potentials, the electrical impulses that enable communication in the brain. Hodgkin and Huxley represented the cell membrane as an electrical circuit, but one feature of their model was initially based on trial and error: a variable controlling the opening of membrane gates had to be raised to the fourth power to match their measurements.
Now loading...
The researchers proposed that this might reflect four charged components moving before a gate opened. Decades later, scientists found that a related potassium channel had four identical parts arranged around a central pore. The model had not been designed to predict that structure, yet it offered a clue to how the channel was built.
The episode illustrates why a model’s value is not simply whether it reproduces data, but what it reveals about the system being studied. A model that fits measurements may still be one of many possible explanations. Knowing which alternatives fail can help researchers distinguish a useful scientific insight from a mathematical match.
More than 70 years after its development, the Hodgkin-Huxley model remains a standard account of electrical activity in neurons. Its equations, adjusted with different constants, describe nerve cells across species, from squid to humans. The work also offers a template for scientific modeling: researchers build a model, adjust its parameters to fit evidence, and interpret what its results imply about the real world.
Artificial intelligence is increasingly being used in each of those stages. One area of progress is fitting parameters. A single set of values may reproduce experimental results, but other combinations could work just as well and lead to different predictions. Bayesian inference addresses this uncertainty by estimating a probability distribution across parameter values, rather than identifying only one best fit.
For complex systems, calculating that distribution directly can be prohibitively difficult. Simulation-based inference offers an alternative: researchers run a model repeatedly with varied inputs, then train a neural network to estimate which combinations could have produced observed data. The method can be computationally intensive at first, but the trained network can analyze new measurements quickly. It has been applied in fields including astrophysics, biology and geophysics, as well as in studies of nerve cells.
AI tools are also being developed to propose models, not just tune existing ones. Systems using large language models can write candidate equations or code, compare their outputs with measurements and revise their proposals. But a model that matches the data does not necessarily explain it. Different structures and assumptions can generate similar results, so presenting only one AI-generated answer risks concealing viable alternatives.
Researchers argue that AI should instead help map a range of plausible models, weighing how well each is supported by the evidence and how much complexity it requires. Shared features among the surviving candidates may be strongly supported, while disagreements can show where more data or a new experiment is needed. This approach also helps scientists test how confidently models predict the effects of changing a system.
Interpreting what a model means remains more difficult to automate than building or fitting one. Hodgkin and Huxley’s suggestion about four moving parts gained credibility partly because simpler versions of their equations did not fit as well. Their model also yielded further predictions about nerve-cell firing, including the existence of a voltage threshold. The failed alternatives helped establish which conclusions the evidence could support.
As computing expands the scale of models, scientists may be able to study systems that are too complex for people to describe or understand by reading equations alone. Researchers are already building large-scale models using a complete wiring diagram of the fruit fly brain, which contains more than 100,000 neurons. Such work could help investigate how neural connections shape behavior, though many different models may remain consistent with the available evidence.
In fields such as climate science and particle physics, researchers already compare families of models and report which possibilities remain compatible with observations. The goal is often not to identify one definitive description, but to determine what conclusions hold across plausible alternatives and which experiments could eliminate the most uncertainty.
AI could compress work that once took Huxley weeks into a matter of hours. But speed alone will not produce scientific understanding. To offer insight, systems must show not only which model fits, but which assumptions are supported, which alternatives failed, where uncertainty remains and what questions should be asked next.
