AI Model Tool ⬤ Live on HuggingFace

Trikal

MOSFET Remaining Useful Life

Data · Physics · AI. Tri = three pillars, Kal = time (Sanskrit). A GBR + LSTM ensemble trained on NASA thermal overstress aging data that predicts how much life a Silicon MOSFET has left — from a single measurement, in ~15ms on CPU.

7.3%
GBR VAL MAE
cross-device
15ms
CPU inference
no GPU needed
6,554
training
snapshots
88.7%
temp feature
importance

Three Pillars

Past data, present physics, future AI — fused into a single RUL score.

01 —
Data
PAST
NASA MOSFET Thermal Overstress dataset. 1 rampup device run to failure, 6,554 snapshots. Package temperature rises monotonically from 29°C to 157°C.
✓ Complete
02 —
Physics
PRESENT
PINN + Arrhenius thermal aging model. Bridges Devsim device simulation (gate length, oxide thickness, doping) to real-world lifetime estimates.
⟳ In progress
03 —
AI
FUTURE
GBR + LSTM ensemble. 90/10 weighting on single points, 60/40 on sequences. Val MAE 7.3%. LSTM val_loss 0.016 after 150 epochs.
✓ Complete

Features

Nine features. Two are critical. Only Rds_on_ohm and package_temp_C are required — everything else defaults to 0.

Feature
Unit
Importance
Description
package_temp_C
°C
88.7%
Primary aging signal — rises 29→157°C monotonically
id_rms_A
A
10.7%
RMS drain current — decreases as device capacity degrades
id_peak_A
A
0.4%
Peak drain current during switching transients
sw_energy_proxy
J
0.1%
Switching energy — increases with thermal damage
Rds_on_ohm
Ω
<0.1%
On-resistance — flat until cliff-edge failure at ~20× baseline
rds_jump_flag
0/1
<0.1%
Set to 1 when Rds exceeds 10× baseline — cliff-edge detector

Sanity Checks

Three snapshots from the training device — start of life, mid-life, and the cliff-edge failure moment.

Start of life — snap 0
T=29°C · Rds=6Ω · id_rms=3.5A
RUL 85.7%
Healthy
Mid-life — snap 3277
T=119°C · Rds=3.15Ω · id_rms=2.5A
RUL 30.4%
Degraded
Cliff-edge failure — snap 6537
T=136°C · Rds=147Ω · rds_jump_flag=1
RUL 0.0%
Replace

Quick Start

Live at xkalp-trikal.hf.space. POST a measurement, get RUL back.

curl -X POST https://xkalp-trikal.hf.space/predict \
  -H "Content-Type: application/json" \
  -d '{
    "Rds_on_ohm":     6.0,
    "package_temp_C": 29.4,
    "id_rms_A":       3.5
  }'

# Response:
{
  "rul_percent":      85.7,
  "health_label":    "Healthy",
  "gbr_rul_percent": 85.6,
  "lstm_rul_percent":86.5,
  "models_agree":    true,
  "rul_hours_est":   685.8
}

Known Limitations

Single training device
Trained on one rampup device. Cross-device R²=0.106. Cannot generalise to mechanical fatigue or gate oxide breakdown without additional data.
Temperature as proxy
88.7% importance on package_temp_C means the model primarily tracks thermal degradation. Devices failing at constant temperature are not well-covered.
Low-temperature blind spot
One Test_N device failed at ~30°C (pred=85%, actual=0%). The model never saw low-temp failure in training. Gap closes with more data.
LSTM single-point divergence
LSTM gets a padded window for single-point queries — it's unreliable there. Use /predict/batch or CSV upload for sequence-aware predictions.

Roadmap

v0.1
NASA MOSFET RUL — GBR + LSTM, Flask API, web UI, HuggingFace deployment
✓ Done
v0.2
Module 1 — Devsim parameter sweep (gate length, oxide thickness, doping)
In progress
v0.3
Module 3 — Arrhenius bridge (simulation → field lifetime)
Pending
v0.4
GaN HEMT support
Pending
v1.0
Multi-device fleet monitoring dashboard
Pending