vsquaretech/ajp-ml-flex-portal

BOND-AI Reliability Digital Twin Portal

BOND-AI is a physics-of-failure digital twin with machine-learning prognostics for high-temperature printed interfaces and bond joints. This FastAPI web app is the reviewer-accessible reference implementation of the standardized qualification methodology proposed under SEMI FlexTech 2026 - Topic B: Advanced Bonding Reliability (UMass Lowell / Raytheon-UMass Lowell Research Institute). It is built around a single 500C-capable alumina qualification coupon that co-locates Zone A printed-interface/interconnect structures with Zone B representative bond interfaces - the attachment layer under study - generates a large synthetic reliability dataset, trains surrogate models, and serves an operator-facing Interface Digital Twin that links degradation of printed conductive inks to bond-interface reliability.

The organizing premise: because every structure shares one substrate, fabrication process, and thermal history, degradation of printed conductive inks can be correlated directly with the electrical and mechanical reliability of the representative bond interfaces they support.

Synthetic demonstrator. All current outputs are illustrative. The five models are trained on 240,000 synthetic rows to demonstrate the data-fusion, conformal-uncertainty, and decision pipeline - not the measured reliability of any specific material. Under the project (Tasks T1-T6) the synthetic training data are replaced with measured coupon data, the q90 bands are re-derived from conformal calibration on observed high-temperature failures, and the decision states are validated against SEM/cross-section ground truth.

The proposal document is not committed. This repository encodes the derived public-facing methodology: qualification structures, inks, representative stress conditions, characterization outputs, reliability targets, and feedback-loop behavior.

Five linked views

Qualification Platform

• —Substrate: single 500C-capable alumina qualification coupon.
• —Inks: baseline Ag-NP conductive ink and 500C-capable ANI conductive ink.
• —Zone A printed-interface / interconnect structures: straight lines (4-point resistance), meander lines (fatigue / crack initiation via dR/R0), and overlap pads (interface degradation).
• —Zone B representative bond interfaces: bond-interface coupons - metallized alumina coupons (~3 mm x 3 mm) attached to printed ANI bond pads with a commercially available high-temperature conductive attachment material - plus daisy-chain continuity structures for non-destructive monitoring of progressive bond degradation and open-circuit detection.
• —Characterization: continuous electrical monitoring, 4-point sheet resistance, bond-interface resistance, daisy-chain continuity, optical microscopy, X-ray/CT, SEM/EDS, adhesion testing (printed interfaces), and post-aging shear testing (bond-interface coupons).
• —Temperature regimes: Regime I (25-180C) - isothermal aging plus RT-180C thermal cycling for both inks; Regime II (250-500C) - isothermal furnace aging at 250, 400, and 500C for the ANI ink only.
• —Experimental scope: approximately 32-36 aging coupons plus 8-10 independent held-out validation coupons; five standardized structures with >= 3 internal replicates each.
• —Independent functional demonstration: a held-out X-band CPW built on the same materials but excluded from model development.

End-of-project performance targets

These are the proposal targets surfaced in the portal's Project Objectives view (objectives contingent on relevant-environment validation, not demonstrated results):

• —Remaining-useful-life prediction: MAPE <= 20%
• —Independent functional demonstration (CPW): correct degradation trend, dominant failure mechanism, and RF performance within model uncertainty
• —Failure-mode classification: Top-1 >= 75%, Top-3 >= 90%
• —Electrical degradation tracking: dR/R0 prediction error <= 15%
• —Active-learning efficiency: >= 25-30% reduction in required coupons
• —Reference material systems: two standardized benchmark systems
• —Measurement repeatability (Gage R&R): <= 10%
• —Standardized qualification methodology: validated for 500C-capable printed conductive inks and representative bond interfaces (TRL/MRL 4 -> 5)

What It Demonstrates

• —Synthetic data generator for process settings, coupon geometry, isothermal-aging and RT-180C cycling exposure, attachment-layer voiding, oxidation, roughness, alignment, drift, delamination, adhesion, and shear.
• —Process characterization model for line width, thickness, resistance, quality, and process anomaly probability.
• —CPW surrogate for the independent functional demonstration on the same alumina material/process family.
• —Coupon reliability model for sheet and bond-interface resistance drift, crack probability, delamination, adhesion strength, attachment voiding, post-aging shear strength, reliability score, and failure mode (including continuity loss and bond degradation).
• —BOND-AI decision layer with PASS, MARGINAL, or INSPECTION_REQUIRED output, failure-mechanism ranking, degradation-state index, confidence, and remaining-useful-life estimate.
• —Interface Digital Twin feedback loop that ranks better process/cure recipes and recommends high-information next coupon tests for active learning.

Local Setup

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
python scripts/train_models.py --pattern-samples 80000 --interface-samples 40000 --coupon-samples 120000
uvicorn app.main:app --reload

Open <http://127.0.0.1:8000>.

Hugging Face Spaces Deployment

For the full ML-enabled main branch, use a Hugging Face Space with Docker.

• —Space SDK: Docker
• —Hardware: CPU basic
• —App port: 7860
• —Branch: main

The Dockerfile pins Python to 3.11-slim, forces binary wheels for scientific packages, installs requirements.txt, and starts:

uvicorn app.main:app --host 0.0.0.0 --port ${PORT:-7860} --proxy-headers --forwarded-allow-ips='*'

No training step is required during deploy when model_artifacts/model_bundle.joblib is committed.

Render Note

Render can still serve this repository when it uses Python 3.11 and binary wheels, but the free build path is slower and more fragile for SciPy/scikit-learn stacks. Hugging Face Spaces with Docker remains the preferred free deployment target for the main branch.

Data Notes

The default training run generates 240,000 synthetic rows: 80,000 print-characterization rows, 40,000 blind CPW validation rows, and 120,000 integrated-coupon rows. The committed data/synthetic_preview.json, data/synthetic_summary.json, and model_artifacts/metrics.json document the generated distributions, target ranges, model benchmarks, and Digital Twin reliability metrics.

To recreate the full local parquet datasets:

python scripts/train_models.py --write-full-data

Full parquet files are ignored by git to keep the deployable repository lightweight.