Trusty Neurocoder¶
Neuro-Symbolic Agents for Verified Scientific Code Generation.
Overview¶
Trusty Neurocoder combines LLM-based agentic workflows with Neuro-Symbolic Abstract Machines (NSAMs) to enable verified scientific code generation, optimization, and surrogate construction.
NSAMs are neural networks structurally equivalent to programming language interpreters. They enable principled compilation of symbolic programs into neural architectures and decompilation back to interpretable code. LLM agents bridge the gap between real-world scientific codebases and the declarative representations NSAMs require.
Notebooks¶
Foundations¶
| Notebook | Description |
|---|---|
| 01 - Cajal Intro | Boolean functions compile to matrices; iteration as recurrent neuron |
| 02 - Exponential Decay | Learn scalar ODE rate constant from data |
Earth & Environment¶
| Notebook | Description |
|---|---|
| 03 - Unknown Function | MLP learns unknown moisture response; symbolic regression recovers Hill equation |
| 04 - CENTURY-Lite | 3-pool model, 2 unknown functions learned simultaneously, mass conservation verified |
DOE Science Domains¶
| Notebook | Description |
|---|---|
| 05 - Decay Chain | 4-isotope radioactive decay chain; learns unknown branching ratios exactly |
| 06 - Battery Degradation | SEI growth + capacity fade; recovers parabolic growth law via symbolic regression |
| 07 - Chemical Kinetics | Reversible reaction A⇌B; recovers Arrhenius rate k=2.0·exp(-5.0/T) from equilibrium data |
| 08 - EcoSIM Kernel Case Study | Soil-carbon decomposition kernel extracted from EcoSIM Fortran |
| 09 - PFLOTRAN Relative Permeability | Subsurface-flow constitutive relation as a structured surrogate |
| 10 - Methionine Cycle | Regulated metabolic cycle (DTU Biosustain model); keeps stoichiometry exact, recovers SAM→CBS allosteric Hill activation, conservation to 1e-7 |
| 11 - Regulated Steady State | Amortizes the regulated steady-state solve; exact moiety reduction (left null space of S), scales to 128 metabolites, ~95x surrogate speedup with conservation guaranteed |
| 12 - Methionine Steady-State Fit | Real methionine+folate network; finds the folate moiety from S's null space; differentiates through the steady state (implicit function theorem) to fit SAM→CBS / SAM→MTHFR allosteric constants from steady-state data |
| 13 - Methionine Bayesian UQ | Bayesian calibration with uncertainty, priors anchored to the real Maud methionine model; Gauss–Newton Laplace posterior surfaces the amp/Ka identifiability ridge (corr 0.99) that the point fit hid |
| 14 - CBS MWC: Laplace vs HMC | Faithful single-enzyme port (real Maud MWC rate law + priors, mM units); Laplace matches exact grid and HMC on informative data, and visibly misses the skewed L tail on data-poor designs |
| 15 - Alzheimer's Aβ–Tau Core | Compiles an SBML/Antimony-style AD reaction model (Aβ aggregation + tau phosphorylation) to a Cajal program; recovers the hidden amyloid→tau hyperphosphorylation coupling as Michaelis–Menten with both conservation laws (total Aβ, total tau) exact by construction |
Quick Start¶
# Install
uv pip install -e ".[notebooks,docs]"
# Run all notebooks
just notebooks
# Serve docs locally
just docs
Architecture¶
┌──────────────────────────────────────┐
│ Layer 1: LLM Agent │
│ - Parses real scientific code │
│ - Extracts algorithmic kernels │
│ - Translates to Cajal programs │
├──────────────────────────────────────┤
│ Layer 2: NSAM Compilation │
│ - Cajal program → PyTorch RNN │
│ - Learnable sub-expressions (MLPs) │
│ - Backprop through compiled program │
├──────────────────────────────────────┤
│ Layer 3: Verification │
│ - Structural invariants by design │
│ - Symbolic regression (decompile) │
│ - Mass conservation, positivity │
└──────────────────────────────────────┘