Physics Validation
Back-test of every model against public HJT industry data, with coupling, uncertainty and inverse solver.
0/5 models passavg R² 0.226
M26M33M35M38M45M62M96M97M99
PECVD i-layer → implied Voc
FAILR²
0.504
min 0.92
RMSE
16.46 mV
max 12
MAPE
1.5%
N = 11
Adachi 2015; Descoeudres 2013; Yoshikawa 2017
Single-diode J0 → Voc at Jsc=39 mA/cm²
FAILR²
0.154
min 0.97
RMSE
18.47 mV
max 8
MAPE
2.5%
N = 6
Green 2013 'Solar Cell Fill Factors'; ITRPV 2024
Kamal-Sourour EVA cure at 600 s
FAILR²
0.822
min 0.93
RMSE
0.095 α
max 0.08
MAPE
15.4%
N = 6
Kempe 2007 (NREL); Hirschl 2013
Microcrack probability vs ribbon tension
FAILR²
0.045
min 0.9
RMSE
0.078 P(crack)
max 0.04
MAPE
75.1%
N = 7
Beinert 2019 (Fraunhofer ISE); Kajari-Schroeder 2012
Bearing BPFO amplitude vs health decay
FAILR²
-0.393
min 0.95
RMSE
2.52 g (norm)
max 0.5
MAPE
73.2%
N = 6
RKB Bearing Handbook accelerated-life rig
Normalized residuals (all models)
Dashed lines = ±5 % envelope. Most points should sit inside.
Monte-Carlo uncertainty (N=400)
95 % CI| KPI | Mean | σ | 95 % CI |
|---|---|---|---|
| Efficiency η | 23.23% | ±0.09% | [23.07%, 23.39%] |
| Voc | 738.6 mV | ±0.3 mV | [737.9 mV, 739.1 mV] |
| Pmax | 6.37 W | ±0.02 W | [6.33 W, 6.42 W] |
| Fill Factor | 83.29% | ±0.04% | [83.21%, 83.37%] |
| iVoc | 739.6 mV | ±0.1 mV | [739.5 mV, 739.6 mV] |
| J0 | 12.45 fA/cm² | ±0.15 fA/cm² | [12.25 fA/cm², 12.81 fA/cm²] |
Inputs perturbed: i-layer ±5 %, T ±1.5 %, RF ±8 %, H₂ ±6 %, doping ±3 %, TCO scan ±3 %, finger width ±6 %, wafer ±6 %, tension ±8 %.
End-to-end coupling: recipe → economics
Δ Pmax
+2.00 W
Δ Yield
+37.4%
Δ Fleet NPV
$-162.48 M
100 MWdc, 25 yr
Δ Warranty / 1k
+1.6
Variant: i-layer 4.5 nm (thin), ribbon tension 5.5 N. Walks the full chain: cell IV → microcracks → module Pmax → 25-yr LeTID/crack-growth → discounted PPA revenue.
| Cell η | 23.45% |
| Stringer crack P | 3.50% |
| Module Pmax (BOL) | 425.1 W |
| Module Pmax (25 y) | 384.7 W |
| Annual degradation | 0.36%/yr |
| Warranty claims / 1k | 4.8 |
Bayesian root-cause solver
Observed Δ: voc_mV=-12 ff_pct=-0.4
| Candidate | Category | Prior | Posterior |
|---|---|---|---|
| #1PECVD i-layer thinning (epi onset) | PECVD | 17.3% | 97.1% |
| #2H2 dilution out-of-band (low) | PECVD | 3.8% | 2.4% |
| #3PECVD chamber temperature drift +25°C | PECVD | 9.6% | 0.4% |
| #4PECVD RF power excursion | PECVD | 4.8% | 0.1% |
| #5PECVD i-layer overgrowth | PECVD | 5.8% | 0.0% |
| #6TCO target poisoning | TCO | 11.5% | 0.0% |
| #7Screen-print finger spreading | 7.7% | 0.0% | |
| #8Stringer ribbon over-tension | Stringer | 10.6% | 0.0% |
Method: forward-model each candidate fault, score Gaussian likelihood against observed Δ, multiply by ITRPV-2024 prior, normalize.