docs/06_explainability.md

06 — Explainability and mechanistic validation (RQ5)

← Back to 05: Decision framework | README

Covers Phase 9. Every number is generated into outputs/reports/.

RQ5: do the learned attributions recover known electrochemical degradation signatures, so a process engineer can trust and act on them? Answer: partly — two hypotheses supported cleanly, one supported but for the wrong physical reason, and one not supported at all.

---

1. Hypothesis verdicts

Four falsifiable hypotheses, stated in advance. Agreement was not forced.

	Hypothesis	Verdict
H1	ΔQ(V) variance features rank top-3 by SHAP at budget ≥ 50	SUPPORTED
H2	Resistance features dominate at very low budgets (5–10)	NOT SUPPORTED
H3	Thermal SHAP is monotonic in the thermal feature value	SUPPORTED, with a mechanism caveat that changes its meaning
H4	Without recipe descriptors, curve features absorb the attribution	SUPPORTED

H1 — SUPPORTED

A ΔQ(V) variance feature sits in the top 3 at every budget ≥ 50: rank 2 at N=50, rank 1 at N=100, where it carries 27.2% of total attribution.

This is the expected result and it is consistent with Gate 1, which reproduced the published log₁₀ var(ΔQ(V)) relationship at R² = 0.859. The model is leaning hardest on the signal the literature says it should.

H2 — NOT SUPPORTED

The prediction was that resistance features would carry disproportionate weight at 5–10 cycles, consistent with formation-stage diagnostics. They do not:

	Resistance share of total attribution
Budgets 5–10	5.3%
Budgets 50–100	7.1%

The share is lower early, not higher — the opposite of the prediction. Two readings are possible and this project cannot distinguish them: either the resistance measurement in this corpus is too coarse to carry formation-stage information (it is a single per-cycle scalar, not an impedance spectrum), or the curve-shape features already capture that information more efficiently at every budget. Reported as a clean negative.

H3 — SUPPORTED, but the mechanism is not thermal

All 4 of 4 thermal features show |Spearman| > 0.7 between feature value and SHAP contribution: the model uses them consistently and directionally, which is what the hypothesis stated.

But the physical reading is wrong, and the check that found this was run because the result contradicted an earlier measurement. Phase 3 measured thermal signal as weak (|ρ| ≈ 0.22–0.25); Phase 9 found thermal_exposure carrying 23.2% of attribution as the #2 feature. That tension was investigated rather than explained away:

Quantity	Spearman vs log₁₀ cycle life	Spread across cells
thermal_exposure (∫T dt)	−0.748	2.56×
temp_mean (per-cycle mean temperature)	+0.222	1.21×

The two phases measured different quantities. Phase 3's was a per-cycle mean temperature; Phase 4's thermal_exposure integrates temperature over time. In a 30 °C-controlled chamber the temperature term barely varies (1.21× across the whole cohort), so most of the integral's variation is cycle duration, and it correlates 0.414 with mean charge time.

Conclusion: thermal_exposure is predictive as a time-under-load proxy, not as an Arrhenius thermal effect. The hypothesis as stated passes; the physical interpretation it was meant to support does not. This is exactly the failure mode the module docstring warned SHAP cannot detect on its own — SHAP reports what the model used, not whether the mechanism is real — and it took an external consistency check to surface it.

The plain-language audit sheet reflects this: thermal_exposure is described to the engineer as "cumulative time the cell spent under load", with an explicit note that it should not be read as a heating effect.

H4 — SUPPORTED

Removing the Group F process-recipe descriptors:

	With recipe	Without recipe
Curve-feature share of attribution	41.1%	47.2%
Cross-validated RMSE (log₁₀)	0.0597	0.0597

Curve features absorb the attribution and the information — RMSE is unchanged to four decimal places over 10 grouped folds. This is the favourable outcome for RQ4: the model is not merely memorising a recipe-to-lifetime lookup, because removing the recipe costs it nothing.

The failure mode this hypothesis exists to detect would have been a rise in curve attribution with degraded performance — attribution moving without information moving. That did not happen.

2. Does the anomalous cell dominate the attributions?

Phase 4 documented b1c41's dq_kurtosis at 112.9 against a cohort IQR of [−1.10, −0.30] — roughly 140 robust-z from the median. A single cell that extreme can dominate a global ranking, so every global result was computed with and without it, and it was never silently excluded.

It does not dominate. At budget 100, removing b1c41:

• the top 7 features are unchanged in rank
• the maximum rank shift anywhere in the top 10 is 2 places
• the largest share change is 0.011 (ir_at_cycle_2, 0.055 → 0.044)

The concern was reasonable and the measurement settles it. Both rankings are in shap_anomaly_comparison.csv.

3. Global attribution at the reference budget

Rank	Feature	Share
1	dq_var	27.2%
2	thermal_exposure	23.2% (read as time-under-load — see H3)
3	q_discharge_curve_area_N	12.2%
4	cc_time_fraction	8.4%
5	recipe_stress_x_thermal	6.7%
6	ir_at_cycle_2	5.5%
7	policy_c_mean	4.4%

Figures: fig18 (global importance), fig19 (beeswarm), fig20 (importance vs budget), fig21 (per-cell waterfalls), fig22 (verdict summary).

4. The QC audit sheet

The artifact that makes this work concrete rather than decorative: a one-page sheet per cell, in outputs/reports/audit_sheets/.

Each sheet carries the prediction and its conformal interval, the assigned grade, the expected cost of every available action (so the reader sees what the alternative would have cost rather than being told the answer), the top-5 drivers in plain language, and any data-quality flags inherited from Phase 2.

A defect found and fixed while building it. The first version quoted feature values straight from the model pipeline — which ends in a StandardScaler, so the sheet was showing z-scores: "how much the discharge curve's shape changed: −0.8224". That is meaningless to the exact reader the artifact exists for. Sheets now quote the physical value alongside the cohort median, so the engineer sees "this cell: 0.000433 (typical cell: 0.00017 — this one is above typical)".

An extract:

     1. How much the discharge curve's shape changed between the early
        and late reference cycles
        this cell: 0.000433   (typical cell: 0.0001696 -- above typical)
        effect on the decision: LOWERED the predicted cycle life
        for reference, a HIGH value means a large change, which indicates
        faster loss of usable lithium

5. Limitations

• SHAP explains what the model used, not whether the mechanism is real. H3 is the concrete demonstration: a hypothesis can pass while the physical story behind it is wrong. Every verdict here is evidence about consistency, not proof of mechanism.
• The audit sheet's usefulness has not been tested with a process engineer. The tests check completeness and internal consistency, which is necessary and not sufficient. Only a human reader can judge whether the explanation lands.
• Attribution is computed on a model fitted to the full cohort, which is correct for "what did the deployed model use" and would be wrong for any performance claim. No performance number is derived from that fit.

---

← Decision framework · README · Protocol shift →