Skip to content

Latest commit

 

History

History
133 lines (107 loc) · 7.13 KB

File metadata and controls

133 lines (107 loc) · 7.13 KB

03 — Milestones and roadmap

Milestones are deliberately small and each ends in something checkable. The loop is the constant; the milestones are what we point it at.

M0 — Hydrogen abstraction (reproduce a known step) · in progress

Claim to test: the loop, unattended, confirms that an ethynyl radical tooltip (H–C≡C·) abstracting a hydrogen from a hydrogenated carbon workpiece is thermodynamically favorable, with the right sign and roughly the right magnitude.

Why this step: it is the textbook first move of mechanosynthesis and its answer is known independently of any simulation — the C–H bond formed in acetylene (H–C≡C–H, ~132 kcal/mol) is much stronger than the C–H bond broken in a typical hydrocarbon (~96–105 kcal/mol), so abstraction should release on the order of 30–40 kcal/mol. A loop that can't recover this shouldn't be trusted on anything novel. (Sign and order of magnitude are the bar; exact numbers depend on method and workpiece.)

Workpiece ladder (cheap → realistic):

  1. methane CH4 (primary C–H) — smallest possible smoke test of the pipeline.
  2. isobutane (CH3)3C–H (tertiary C–H) — the chemically relevant weak C–H.
  3. adamantane C–H — a rigid cage that better models a diamond surface site.

Done when: the ledger contains, for at least the isobutane case, a reaction-energy record with ΔE < 0 of the expected order, produced by the loop without hand-holding, and the journal's M0 entry reports the number with its method and caveats. That satisfies project criterion S1.

M1 — Feasibility, not just favorability · in progress

Add the approach-coordinate scan to the arbiter so survivors report a barrier under mechanical approach, not only a reaction energy. Distinguish steps that are downhill-but-blocked from steps that actually proceed under positional load.

Status: cheiron.approach builds the collinear C–H···Tool· supersystem at a controlled approach distance and runs both a rigid scan and a constrained relaxed scan (geomeTRIC $freeze on the approach distance, optimized geometries recorded per point, references audited against the M0 ledger). Production def2-SVP results: ethynyl+methane barrier 0.0 kcal/mol, methyl+methane ≈8.2 — the feasibility discrimination works. score() now takes the barrier: feasible gate at 15 kcal/mol, fitness = −ΔE − 2·barrier (declared heuristic). Remaining for M1: scan the tool/workpiece shortlist (ethynyl+isobutane first), surface barriers in the published summary, and a hybrid-functional cross-check of the PBE barriers.

M2 — Selectivity and tool integrity · complete

Introduce workpieces with multiple inequivalent C–H sites and score whether the tool geometry selects the intended one. Add the tool-integrity check as a hard gate. First candidates for a novel datasheet (criterion S2) come from here.

Status: site-selectivity measurement complete for ethynyl on adamantane. Thermodynamic margin: tertiary preferred by 1.14 kcal/mol (tool-independent). Kinetic margin under clamped-bodies positional control: zero — both sites barrierless, entrance wells 0.2 kcal/mol apart, transfer complete by 1.8 Å. Conclusion (the M2 headline): on diamondoid workpieces, site selectivity is entirely positional; chemistry does not discriminate. Method notes: one-leash scans mismeasure crowded sites (tool slides off-target — measured); the clamped-bodies scan (two Cartesian anchors per body, transferring H free) is the positional-control model. Remaining for M2: the tool-integrity hard gate, and a hybrid-functional spot-check of the PBE numbers.

M3 — Search, not enumeration · in progress

Turn on the evolutionary and agent proposers; let the loop explore the tool/workpiece space rather than a hand-written grid, seeded by everything the ledger has learned. Publish the negative results this generates (criterion S3).

Status: the tool space is widened (hydroxyl/amino/vinyl join ethynyl/methyl, spanning the abstractor-strength axis) and cheiron.predict fits the additive ΔE model from the ledger (worst residual on the 16-record training set: 0.000 kcal/mol) and ranks the unevaluated frontier. Strategy: one anchor measurement per new tool (× methane) pins that tool's whole ladder; the model proposes, the arbiter disposes, and predicted-vs-measured deltas get published either way. PBE0 spot-check bounded the method bias (barriers are lower bounds, ≈ +2–3 kcal/mol at hybrid).

M4 — A second operation: radical addition · in progress

Prove the loop is not hardwired to hydrogen abstraction by taking a bond-forming step — a radical adding across a C=C — Tool· + alkene → Tool-CH2-CH2· — through the same PROPOSE→BUILD→ARBITER→SCORE machinery and append-only-ledger discipline.

Status: the operation is live and validated. cheiron.addition builds the three-species system (isolated from the abstraction pipeline, which stays green); experiments/m1_radical_addition/run_addition.py scores ΔE via the existing arbiter. Validated against known chemistry (methyl + ethylene −32.1 vs literature electronic ΔE ≈ −28…−30; PBE overbinds a few kcal as it does for abstraction). Two findings the single-operation loop could not have produced:

  1. Tool ranking is operation-dependent — abstraction strength does not predict addition strength (Spearman ρ ≈ 0.8, but methyl is the worst abstractor and a solid adder; hydroxyl/vinyl swap). SELECT must rank tools per operation.
  2. Addition ΔE is approximately additive (tool-difference substrate-independent to ~0.7 kcal/mol across ethylene/propene) — but, unlike abstraction's exact Hess-law additivity (0.0001 kcal/mol), it is empirical and must be validated, not trusted.

Remaining for M4: an approach-barrier adapter for the bond-forming coordinate (the M1 machinery is abstraction-shaped), a SCORE path, and anti-Markovnikov site selectivity on substituted alkenes.

Cross-cutting: continuous operation

Wire the harness to a scheduler so a batch runs, the ledger and journal update, and any human requests surface — on a cadence, across sessions, without a person driving each step. The in-process loop already runs unattended for a bounded run; this makes it perpetual.


External help wanted

Collected asks for the human (and, through them, other people/systems). None block M0.

  • A working GFN2-xTB (conda xtb/tblite, or a compiled binary on PATH). Gives the loop a ~100× cheaper screening tier and access to much larger workpieces than DFT allows on this host. Currently blocked by: no sudo/conda on the build host.
  • A modest higher-accuracy compute budget for Tier-3 confirmation runs (larger basis / hybrid functionals / dispersion) on shortlisted candidates.
  • A domain expert to hold the VETO pen — someone from the molecular nanotechnology / computational chemistry community willing to review promoted candidates and reject bad directions. The gate is real but currently unstaffed.
  • Pointers to reference data — published DFT energetics for specific mechanosynthesis steps we can calibrate the arbiter against.