Chemistry Lab: Exploring the Limits of AI-Built Scientific Simulations

Try the interactive demo: Chemistry Lab

Chemistry Lab continues our experimentation with GPT Astra through an interactive environment for exploring atoms and molecules. The project examines how AI-assisted development can turn scientific concepts into usable software, and why reference data, explicit assumptions, and human guidance remain essential.

Following our Newton Lab experiment, Chemistry Lab brings the investigation into atomic structure and chemical bonding. Its central question is how faithfully an AI-assisted application can represent scientific relationships while making them accessible through interaction.

This demonstrator introduces several overlapping layers of complexity. Electronic configurations, nuclear stability, and molecular bonding require different models. A useful explanation at one level may be insufficient at another. Chemistry Lab makes these boundaries visible through two connected activities: Atom Builder and Molecule Builder.

Atom Builder is the main experiment. Users change the numbers of protons, neutrons, and electrons, then inspect the resulting element, isotope, charge, and electronic configuration. The workbench covers all 118 elements, with an interactive periodic table, shell and orbital views, and guided challenges.

Changing one particle can alter the meaning of the entire configuration. Adding a proton changes the element. Adding a neutron changes its isotope. Adding or removing an electron changes its charge. Bringing these controls together lets users explore relationships that are often introduced separately in textbooks.

The application draws on bundled reference data to distinguish stable isotopes, radioactive isotopes, and configurations for which information is unavailable. That third category matters: missing data cannot justify an invented prediction. Likewise, familiar electron-filling rules need exceptions and carefully identified estimates. A consistent-looking diagram can still conceal an incorrect scientific assumption.

Another important distinction concerns the word “unstable.” A radioactive nucleus and an atom that cannot retain an additional electron involve different processes. Atom Builder treats nuclear stability separately from electron binding.

In its isolated-atom mode, a simplified model uses reference electron affinities to determine whether excess electrons should remain bound. Construction mode retains user-selected electron counts for exploration. These modes expose the assumptions behind the behavior: an isolated atom and an ion considered within a chemical environment need different treatment.

The visualizations also have defined limits. Shell diagrams are schematic, and electron-escape animations illustrate the model’s outcome. They do not calculate individual electron trajectories, escape times, or quantum wavefunctions.

One of the clearest demonstrations is beta-minus decay. For supported radioactive isotopes, users can manually step through a nuclear transformation. Carbon-14 provides a concrete example: one neutron becomes a proton, with the emission of an electron and an electron antineutrino. The nucleus becomes nitrogen-14. Its proton number increases from six to seven, while its mass number remains fourteen. This transformation follows the standard beta-decay description provided by NNDC.

The simulation keeps the bound electron count unchanged during this step. Starting with neutral carbon-14 therefore produces an immediately positive nitrogen ion in the model. The emitted beta electron carries away the balancing negative charge; it is not added to the atom’s electron shells.

This creates a useful connection between the animation and conservation rules. The interface must account for what changes, what remains constant, and where the emitted particles belong. Those relationships provide stronger checks on correctness than visual plausibility alone.

The manual control represents a supported decay channel. Actual radioactive decay occurs spontaneously, and the animation’s duration has no relationship to the isotope’s half-life. The current implementation does not simulate competing decay probabilities, energy sharing, nuclear recoil, or excited daughter states.

Those omissions point toward the next scientific challenge: representing energy states explicitly.

A more advanced laboratory could introduce irradiation, allowing users to explore interactions with incoming photons or particles. Such an extension would need to specify the incident radiation, its energy, and the interaction being modeled. An “irradiate” button would require a physical model capable of explaining the resulting changes.

Electronic excitation and nuclear excitation would also need separate treatment. An electron changing its energy state concerns the atom’s electronic structure. Nuclear excited states concern the nucleus itself, including longer-lived excited configurations known as isomers. Gamma emission can lower nuclear energy without changing the numbers of protons or neutrons, as described in Berkeley Lab’s introduction to gamma decay.

Beta-plus decay introduces another distinction. It converts a proton within a suitable nucleus into a neutron, emitting a positron and an electron neutrino. It requires a favorable nuclear mass-energy balance; lowering the atom’s electronic energy does not trigger it. Electron capture also converts a proton into a neutron, but consumes an atomic electron and emits a neutrino. Its energy requirements differ from positron emission. These mechanisms would need explicit treatment in any extension, following evaluated nuclear-decay principles.

Molecule Builder adds a complementary perspective. Users connect atoms, adjust bond orders, inspect electron accounting, and work through guided challenges using editable Lewis structures. This shifts attention from the composition of an individual atom to the distribution of electrons in bonds and lone pairs.

Here, too, the interpretation has limits. A valid Lewis structure demonstrates consistent electron accounting within the selected rules. It does not establish that a molecule is physically stable, predict its three-dimensional geometry, or determine how it will react. The diagram supports a specific kind of reasoning.

Together, these experiments make Chemistry Lab a useful setting for examining AI-assisted scientific software development. GPT Astra’s role in producing the application belongs within a broader process of selecting references, defining models, checking results, and reviewing explanations. A working interface is one part of that process.

The repository includes checks for electronic configurations, isotope classification, charge and particle accounting, and supported decay behavior. These checks help establish whether the software follows its stated rules. Scientific review must also examine whether those rules are appropriate and whether users can understand their limits.

Chemistry Lab remains an educational demonstrator. Its educational effectiveness has not yet been established through learner studies. Further development can deepen both the science and its evaluation: adding carefully bounded energy-state models, testing explanations with educators, and investigating how learners interpret the simulations. Each extension should make the laboratory’s behavior more informative while keeping the evidence and assumptions open to inspection.