A Reduction and a Dipeptide Through the ChemIllusion ELN
A walkthrough of the ChemIllusion lab notebook with two examples. A sodium borohydride reduction shows the record itself: scheme, hands-free notes, checks and write-up. A dipeptide, L-alanyl-L-alanine, shows the chemistry tools: processing its ¹H NMR, asking the Structure Solver what it is, and predicting the mechanism of its hydrolysis.
Updated October 18, 2026: new screenshots of the current notebook, and a second worked example with a real dipeptide spectrum. The September version described features; this one shows them.
The two examples
The easiest way to explain the notebook is to run something through it, so this post uses two examples.
- The record: sodium borohydride reduction of acetophenone to 1-phenylethanol in methanol. Most of us did it as undergraduates. One reactant, one reagent, one product. It is a demo record, so the amounts are made up, but every screen is the real interface.
- The chemistry tools: L-alanyl-L-alanine (Ala-Ala). I used a real, public ¹H and ¹³C data set for it, and every number in that part of the post came out of our own processing code and Structure Solver.
1. Open the notebook
Rail on the left, your records in the middle, the work on the right. Each row shows the product, so you can find a run by its structure and not only its title.
Type a title in New experiment and click Create draft. If the run belongs to a project, file it there. The “Worth a look” feed points out loose ends across your own records, such as a product you isolated but never analyzed. You can dismiss anything in it, and nothing in it changes your data.
2. Draw or say the reaction
The Scheme tab. The reaction is stored as real chemistry (reaction SMILES), and the stoichiometry table pulls molecular weights and densities from the materials library.
You can get the reaction in three ways:
- Draw it with Edit in Ketcher.
- Type or paste a reaction SMILES or names.
- Scan structure from a photo of a drawn structure.
Pick the reference reactant and product. The notebook uses that pair to find your other runs of the same step later. It never picks them for you without telling you.
The stoichiometry table keeps planned and actual amounts in separate columns. When something is missing, it tells you exactly what. In this run I hadn't entered purities, so it says it is assuming 100%.
The Assistant on the right is labeled Proposals only. Nothing is recorded without you. Its chips (Check my stoichiometry,Compare with my previous runs, What should I expect to see?) are shortcuts. Quantities and yields come from the notebook's own calculations, not from the model.
Saying it instead. If you just want a structure on screen, theSpeak tile in MyMol, our phone and tablet app, does that from a spoken name. The Build It, Say It quick guide covers that flow. Inside the ELN, the same idea lives in the hands-free helper in the next step.
MyMol’s start screen on a phone. Speak is the third tile.
3. Hands-free at the bench
The full command list is on screen, so nobody has to memorize it.
When I'm wearing gloves, I don't want to touch the screen. Put the tablet on a stand, open the run, and tap Start hands-free once. After that it is voice only:
- Say “take notes” and then talk. “Added NaBH₄ in three portions over five minutes, flask stayed below 5 °C.” Notes are cleaned up for chemical names and units, and they stay drafts until you say“save notes”.
- “voice draw” lets spoken commands edit the drawing, and “molecule four is the product” sets a role.
- “chat draw” opens a spoken conversation with the assistant about this experiment.
- “go to sleep” when you're busy, and “hey notebook” to wake it up.
On a tablet the molecules are listed under the drawing. Each role you set, by voice or by tapping, writes back to the reaction.
Two things worth knowing. While it's asleep it is still listening for the wake word. That uses transcription minutes, but nothing else is kept, and it switches itself off after ten minutes asleep. Finalizing or abandoning a run can never be done by voice. The helper is still in beta while we test it on more devices under real lab conditions.
4. Keep the scheme in view
The strip and the progress panel stay at the top while you move between tabs.
The Procedure tab. Steps are typed actions you can reorder, not a block of free text.
On paper the scheme is at the top of the page and you look up at it while writing the procedure. The pinned strip does the same job here: limiting reagent, scale and yield are always visible. The progress panel under it tells you what is done and what is next.
5. Let the notebook check you
The Analysis tab tells you what it checked, what it found, and what it could not check.
None of this blocks you from finalizing. It shows you what a careful labmate would point out. In Record review it asked whether borohydride at 0.5 equivalents is really a reagent or should be a catalyst. It's a reagent, since each BH₄⁻ can deliver up to four hydrides, but the question is fair. It also noted that the product row still has no amount.
I like the Not checked list in Outlier reviewmost. It says it can't call the yield an outlier until there are three earlier runs of this step. A check that didn't run is not reported as a pass.
6. Mechanism and write-up
A predicted mechanism is a proposal. It only becomes part of the record when you accept it, and accepting it writes a new revision.
Predict mechanism proposes a curved-arrow mechanism for the scheme you drew. Each step goes through deterministic chemistry validators and is labeled with the engine that produced it. Section 9 shows what happens before that call, using the dipeptide.
The Write-up tab drafts an experimental paragraph from what you actually recorded. Then Finalize. Finalized runs are never overwritten. If you change something later, it is saved as an amendment in a hash-chained revision history you can verify.
Acetophenone
1-phenylethanol
Starting material and product, rendered from the saved structures with the same CineMol engine MyMol uses for its 3D view.
7. A dipeptide: process the ¹H NMR
The borohydride run is a demo record, so it has no real spectrum. For the NMR tools I wanted real data, and a dipeptide is a good test. Ala-Ala in D₂O gives two methyl doublets and two alpha-CH quartets that sit close together, and the NH and OH protons have exchanged away. The near misses, Ala-Gly and Gly-Ala, are close enough to make a solver work for it.
The data is the public BMRB entry bmse000403: L-alanyl-L-alanine, 500 MHz, D₂O, DSS as reference, water presaturation.
The NMR hub. Experimental NMR is the processor; the Structure Solver lives inside it.
Upload the zipped experiment folder. I uploaded the ¹H and the ¹³C experiment separately.
The processed spectrum. The two alpha-CH quartets are on the left and the two methyl doublets on the right.
The peak list matches the BMRB assignment: quartets centered at 4.13 and 4.01 ppm, and methyl doublets at 1.52 and 1.35 ppm, J = 7.3 Hz. The processing reports itself as high confidence because it agrees with Bruker's own processed spectrum (correlation 0.996).
That agreement took some setup, and I'd rather say so. This data set was recorded in 2007 on XWIN-NMR 2.6, and the uploader currently accepts TopSpin 4 data only. To make these screenshots I ran our processing code on it directly with that version check switched off. Automatic phasing also failed on it, so I set the phase by hand (zero order 105°, first order −152°) and turned baseline correction off. The controls are in the processor, and it records that the phase was set manually. The summary line above the spectrum still says “auto-phased”. That wording is wrong, and the fix is ready: the summary will name the phase source and baseline that were actually used.
Every processing choice is editable, and the record shows which ones were automatic and which were yours.
The features the solver works from. Note what it did with the overlapping signals.
This table is worth reading carefully, because it is what the solver uses:
- The processor groups by signal region. Both alpha-CH quartets became one “multiplet”, and both methyl doublets became one “quartet”.
- The integrals are low for the alpha protons. Six methyl protons against two alpha protons should give 1.00 to 0.33. It measured 1.00 to 0.20. Bruker's own processed spectrum gives about the same, so I think it is the presaturation at 4.70 ppm with a one-second delay, which partly saturates protons near the water signal.
- DSS is labeled “likely TMS”. It is the right peak but the wrong name.
None of this is hidden from you. The experimental text it drafts, ¹H NMR (500 MHz, D₂O) δ 4.05 (m), 1.43 (q, J = 7.3 Hz), 0.00 (s), is honest about what was measured, and you would edit it before it went into a paper.
8. Ask the Structure Solver
I opened Solve the structure with the defaults, and it declined.
First attempt. It would rather stop than fit the wrong compound.
With the low alpha integrals, the 4.05 region looked too small to belong to the main compound, and the solver wondered if it was 2-propanol. I told it the 4.05 signal belongs to the compound and that the 0.00 singlet is the reference. Then it ran.
Reclassified, with the integral reading it settled on and the “start from a guess” panel.
If you don't have a structure, it will offer starting guesses from the spectrum alone. For this spectrum they weren't close (ethanol and some small amines), which is about what you'd expect from two broad regions with no carbon data. They're labeled as guesses to test, not answers.
I already knew what to check, so I used Compare two structureswith Ala-Ala and the closest alternative, Ala-Gly. From the ¹H alone they came out exactly the same: 7 fits, 1 conflict, 3 unresolved each. That is the right answer. Two overlapping regions with suspect integrals can't tell them apart.
So I attached the ¹³C spectrum of the same sample. The processor found six carbon peaks: 182.6 and 173.4 (two carbonyls), 54.0 and 51.8 (two alpha carbons), 19.8 and 19.5 (two methyls). There was also a small peak at 22.1 ppm, about 1.4% of the tallest. I marked it as an impurity.
The ¹³C spectrum, as drawn by the processor’s figure export. The inset is the two methyls, 0.33 ppm apart.
It also suggests the next experiment: a DEPT-135 would separate Ala-Ala from the other two.
With the carbon count in, and Gly-Ala added as a third candidate, they separated. Ala-Ala: 11 fits, 1 conflict, 3 unresolved. Ala-Gly and Gly-Ala: 9 fits, 3 conflicts each, mainly because six carbon environments don't fit a five-carbon molecule.
The ledger for Ala-Ala. Every row says what was observed, what the structure predicts, the result, and why.
The one conflict left on Ala-Ala is the methyl region integral. That is the same low alpha integral from step 7, now showing up as a disagreement instead of being quietly ignored. I would rather have that than a clean score. The solver doesn't say “this is Ala-Ala”. It says Ala-Ala fits best, here is the one thing that doesn't fit, and here is the experiment that would settle it. The solver is deterministic. No language model looks at the spectrum.
For teaching labs, the NMR Lab quick guide shows the same flow with students. They process the data, then have to explain the shifts and integrations instead of being handed the answer.
9. Hydrolysis in the Mechanism Predictor
With the structure settled, the obvious mechanism question for a dipeptide is amide hydrolysis: Ala-Ala plus water gives two alanines.
Before any model runs, the reaction is checked: structures parse, stereochemistry is kept, and atoms and charge balance.
Draw the reaction or type the SMILES. Before anything goes to a model, it runs a preflight, and those checks are deterministic. It confirmed that both stereocenters stayed L, that the equation balances, and that this is a new reaction for me. There was no stored reference mechanism for it to replay, so a prediction would be a new run costing 5 AI Actions.
Then I ran it. For the run I added HCl as the reagent and set pH 1, 100 °C and a limit of 14 steps. Benchmark contribution was off.
How the search went: which idea it picked at each step, what it rejected, and what it kept as an alternative.
It accepted six steps:
- H₃O⁺ protonates the amide carbonyl oxygen (C₆H₁₂N₂O₃ + H₃O⁺ → C₆H₁₃N₂O₃⁺ + H₂O).
- Water attacks the carbonyl carbon and gives the tetrahedral intermediate.
- H₃O⁺ protonates the N-terminal amine. The page marks this as a spectator equilibrium at pH 1.
- Water takes a proton from the added oxygen, which gives a neutral gem-diol.
- The amide nitrogen is protonated.
- The C=O reforms and the C–N bond breaks. The run stops with “Target reached”.
Step 2, the water attack. The electron-change table lists every bond made and broken, and it sums to zero.
Step 2 is the one I'd show first. Two numbered arrows, a before and after drawing, and a table of electron changes on the three atoms involved (C4, O5 and the incoming water oxygen O6) that sums to zero. All four checks pass, and the text under the drawing gives the balanced equation, C₆H₁₃N₂O₃⁺ + H₂O → C₆H₁₅N₂O₄⁺.
Step 4, a proton transfer to water that leaves the neutral gem-diol. The water that was added in step 2 hands its extra proton to a second water.
Step 4 is a good one for students because nothing dramatic happens: a second water (O7) takes a proton from the O6H₂⁺ that was added in step 2, giving H₃O⁺ and the neutral gem-diol. The page writes it as C₆H₁₆N₂O₄²⁺ + H₂O → C₆H₁₅N₂O₄⁺ + H₃O⁺. The charge is still 2+ on the left because the N-terminal amine was protonated in step 3. It is a fair thing to ask a student to explain before showing them the drawing.
The order is the one I'd draw on the board: activate the carbonyl, add water, move protons, make the nitrogen a leaving group, collapse. Steps 3 to 6 each kept one alternative idea, so you can see what else it weighed at those points.
It is not perfect, and these are the things I'd want a student to find on their own:
- Step 1 contradicts itself. The selected idea shows a red ✕ for “Atoms and charge balance (failed)” next to three passed checks. The run summary says every step passed its checks, and Idea 2 for the same step was rejected for failing that same check. The equation it gives, C₆H₁₂N₂O₃ + H₃O⁺ → C₆H₁₃N₂O₃⁺ + H₂O, does balance when I count it. So it is an inconsistency between the check, the summary and the rejection, and you should check the step yourself rather than trust any one of them.
- The final state is not what you'd isolate at pH 1. The last step, the C–N cleavage, gives neutral alanine from the C-terminal residue, and the N-terminal alanine as an O-protonated acid with a protonated amine, a dication (C₆H₁₆N₂O₄²⁺ → C₃H₉NO₂²⁺ + C₃H₇NO₂). At pH 1 both alanines should be cations, each with NH₃⁺ and a neutral COOH. One proton has to move from the protonated acid to the free amine. “Target reached” means the structures match the product I drew, not that the protons are where they'd end up.
- Step 3 is chemically true but early. At pH 1 the N-terminal amine is protonated before anything else happens. Putting it in as a separate step in the middle is a bookkeeping choice, and the page says as much by calling it a spectator.
That is how I'd use it with students: as a draft mechanism they have to check, not an answer key. As in the ELN, it stays a proposal until you accept it.
Where the AI sits
Short version:
- Computed, not generated: structures, stoichiometry, yields, record and outlier review, NMR processing, the Structure Solver, and the mechanism preflight checks.
- Proposed by a model, approved by you: cleaned-up voice notes, assistant answers, predicted mechanisms and draft write-ups.
- Recorded with every AI result: which model and tool produced it, the inputs, and whether you accepted, edited or rejected it.
Pricing and the 14-day trial
Academic seats are $50/month per seat, billed monthly with no annual commitment. That covers the ELN, NMR processing, the Structure Solver and the hands-free helper (beta). Enterprise pricing is on request.
Start a 14-day free trial. No credit card. A person reviews each trial request, usually within a day. Or see ELN plans and pricing.