Method · Explainable by construction

What enters GFIS, what the neural engine learns, and what physics is allowed to change.

The evaluator’s black-box concern is addressed by exposing the model as a chain of inspectable transformations, constraints and evidence—not only a final methane number.

The complete reasoning chain

Process measurements
Past-only temporal features
Tabular + neural predictions
Physics and stability audit
Decision + evidence memory

Inputs with physical meaning

  • Digester and ambient temperature
  • pH and organic loading rate
  • Hydraulic retention time
  • Total and volatile solids
  • C/N ratio and moisture
  • Prior methane observations
  • Reactor identity and time

Inputs GFIS must not invent

  • Missing methane targets are never imputed.
  • Total biogas is never renamed as methane.
  • Validation/test observations never fit scaling.
  • Synthetic perturbations are labelled synthetic.
  • Reactor identity is retained across sequences.

Current hybrid engine

XGBoost branch

Learns nonlinear interactions among present process variables and past-only lag/rolling features. Tree structure can be inspected through feature importance and SHAP in the final explainability layer.

LSTM neural branch

Reads an ordered window of reactor observations and retains temporal state through input, forget and output gates. Every epoch, seed, window and checkpoint is stored.

Physics/stability branch

Audits methane feasibility against a VS-dependent ceiling and interprets VFA/ALK as Stable, Warning or Critical. The current ceiling is post-prediction; a genuine loss-term ablation is the next engine upgrade.

ŷhybrid,t = α·ŷLSTM,t + (1−α)·ŷXGB,t

The ensemble weight α must be selected using validation data—not the final test period. In the public DS-03 experiment, the LSTM was not promoted merely because one final-period score looked favourable.

Inside the neural engine

A neural engine already exists: the PyTorch LSTM. Version 3 makes its state update explicit and defines the next physics-guided neural engine as an evidence-controlled extension, not an unexplained new model.

it = σ(Wxixt + Whiht−1 + bi)
ft = σ(Wxfxt + Whfht−1 + bf)
ct = ft⊙ct−1 + it⊙tanh(Wxgxt + Whght−1 + bg)
ŷt = wyTht + by

Existing neural engine

One-layer LSTM, hidden size 32, chronological windows 3/7/14/28, repeated seeds, early stopping, saved epochs and checkpoints. It is real and executable, but DS-03 evidence does not support longer memory as champion.

Proposed physics-guided residual neural engine

A compact recurrent network predicts a correction around a mechanistic or empirical baseline. The correction is penalized when it violates methane-yield bounds or inconsistent stability states. This must be implemented as an ablation-controlled experiment before being claimed as a result.

ŷCH₄,t = ŷreference,t + fθ(xt−w:t)
Ltotal = Ldata + λyieldLyield-bound + λstateLVFA/ALK + λsmoothLtemporal

Each λ will be configuration-controlled. The report will show the accuracy–consistency trade-off, violation rate and calibration—not only the best error score.

How GFIS removes black-box behaviour

Black-box riskGFIS transparency controlVisible evidence
Unknown data originImmutable archive, licence, citation and SHA-256Dataset catalogue and run manifest
Future leakageChronological splits and fold-local preprocessingSplit manifest and rolling folds
Unexplained temporal memorySequence windows, seed repetition and epoch historiesEpoch CSV, checkpoints and window study
Physically impossible methaneRaw prediction, upper bound, violation flag and correction retainedSimulator trace and physics-ablation plan
Opaque alarmVFA/ALK value, thresholds and stability state shown togetherWarning rows and simulator memory
Single favourable metricPersistence baseline, rolling folds and dispersionFold metrics and scientific decision