Doing Science by Chatting with AI
Open a browser, type a sentence, and watch real molecular science happen. How conversational AI is widening the door to discovery — and where AIMS fits in.
When the cost of trying an experiment drops to the cost of typing a sentence, the number of people who can do science goes up by orders of magnitude.
For most of history, doing science meant earning your way in. You needed a lab, or a supercomputer account, or years of training just to ask a question the right way. The tools of discovery sat behind a high wall, and the price of admission was expertise.
That wall is coming down — and the wrecking ball is a chat box.
At AIMS, you can now open a browser, type a sentence in plain English, and watch real molecular science happen. Ask for a molecule and an AI agent generates it. Ask what happens when you heat it, and a molecular-dynamics simulation runs right in front of you. Ask which candidate binds best to a protein pocket, and you can screen them by docking. No installation, no command line, no PhD required. You describe what you're curious about, and the in silico engine does the heavy lifting.
What this looks like in practice
Here's a real exchange. A user types one sentence; the AI agent clears the workspace, fetches a molecule, runs a short molecular-dynamics trajectory, and turns the result into a spectrum — then explains what the peaks mean and where to be skeptical.
Clear the lab, load benzene from the gallery, and compute its IR spectrum.
Here are the results! I cleared the lab, loaded benzene, and computed its IR spectrum. Here's a summary:
✅ Actions performed
- Cleared lab — removed the quaterphenyl molecule.
- Loaded benzene into the simulation chamber.
- Recorded 2.0 ps of dipole-current trajectory (1005 samples).
- Computed IR spectrum — 3 peaks detected.
📊 IR spectrum peaks (wave numbers & relative intensities)
| Peak (cm⁻¹) | Relative intensity | Consistent with |
|---|---|---|
| 2411 | 1.00 (strongest) | C–H stretching modes — somewhat red-shifted from the typical ~3030 cm⁻¹ range, likely due to the classical force field's harmonic parameters |
| 1694 | 0.40 (moderate) | Aromatic C=C ring stretching (~1600 cm⁻¹) — consistent with the ring-breathing modes of benzene |
| 684 | 0.07 (weak) | Aromatic C–H out-of-plane bending (~670–900 cm⁻¹) — a hallmark of benzene's hydrogen wagging |
Interpretation. The spectrum shows the expected C–H stretching (strongest), C=C ring stretching, and C–H out-of-plane bending signatures characteristic of benzene. The shifts from literature values are typical for classical MD with heuristic force fields — the assignments should be considered consistent with these vibrational modes rather than exact.
⚠ Note: Benzene is a non-polar, symmetric molecule, so its partial charges may be zero or near-zero depending on the force field, which can produce a weak or absent IR signal. The peaks above indicate some partial charges were present in this model.
Notice what the agent does besides the calculation. It reports exactly which actions it took, shows the trajectory length it used, and — crucially — flags where the classical force field bends the answer away from the textbook numbers. That last paragraph is the part that matters most: the tool is teaching you to read its own output critically, not asking you to take it on faith.
The bottleneck was never curiosity
We think this is a bigger shift than it first appears. The bottleneck in science has rarely been curiosity — it's been access. The student who wonders why dry ice forms the crystal it does, the tinkerer who wants to know if a polymer chain will hold, the citizen scientist with a hunch about a drug-like molecule: historically, all of them hit the same wall. The simulation existed, but it spoke a language they didn't.
Conversational AI changes the language of the wall. Instead of learning the software, you talk to it.
Part of a fast-moving field
Turning large language models into scientific collaborators is one of the most active frontiers in the field right now. Systems like ChemCrow and Coscientist showed that an AI agent can plan and even physically execute chemistry; copilots like ChatMol and others have brought natural language to molecular modeling, and a wave of new platforms is racing to make computational chemistry conversational. The direction is real, and it's gathering momentum across academia and industry.
What we care about at AIMS is who gets to come along. Most of these systems are built for experts who already live inside the lab. We're building for everyone standing outside it. That means the science has to be visual — you should see the atoms move, not parse a text file. It has to be immediate — it runs in the browser tab you already have open. And it has to be forgiving — you should be able to ask a naive question and get a real answer, then a better one, without being punished for not knowing the jargon.
This is what we mean by lowering the barrier to the in silico approach. And when more people can poke at the molecular world, more of them will notice something nobody was looking for. That's not just convenience — that's how the funnel of discovery gets wider at the top.
Chat is the on-ramp, not the autopilot
There's a healthy skepticism that belongs here, and we share it: an AI that sounds confident is not the same as an AI that's correct, and a simulation you didn't set up yourself is easy to trust too much. Good conversational science tools have to keep the human in the loop — showing their work, exposing the model and the assumptions, and letting you check the AI against reality.
But the on-ramp matters. A generation of scientists grew up being told that the molecular world was off-limits unless they had the right credentials and the right cluster. We're telling them something different now: you can just ask.
← Back to AIMS