# DL4Adhesion M1-concat model card

## Intended use

This recurrent neural surrogate predicts a 120-step dimensionless force
sequence for a rigid sphere interacting adhesively with a linear viscoelastic
substrate under a loading–dwell–unloading indentation protocol. The public
demonstrator is intended for scientific exploration and communication within
the REAL-ADHERE project (Grant 101285102).

It is not a replacement for a boundary-element simulation, a formal
uncertainty model, safety-critical engineering analysis, or experimental
validation.

## Inputs and output

The user specifies dimensionless loading rate, dwell time, unloading rate,
Tabor parameter, and peak indentation. The application constructs 50 loading,
20 dwell, and 50 unloading measurement points. The neural input is a
standardized 120 × 4 sequence (double-symmetric-log time, symmetric-log
indentation, causal velocity, and transition-edge indicator), concatenated
with a transformed scalar Tabor parameter. The output is a dimensionless force
sequence after target inverse-standardization and inverse symmetric log.

## Architecture

M1-concat uses RepeatVector and Concatenate fusion, two sequence-returning LSTM
layers with 256 and 128 units, and TimeDistributed dense layers with 64 and 1
units. The original HDF5 model and a TensorFlow.js Layers conversion are
included as provenance artifacts.

## Domain qualification

The status display combines a five-nearest-neighbour score in robust-scaled
five-parameter space with stepwise training envelopes for time, indentation,
and causal velocity. Thresholds are the 95th and 99th percentiles of
leave-one-out training scores. This is an empirical support indicator—not a
calibrated uncertainty, confidence interval, or probability of accuracy.

Predictions outside the training support are still computed when the protocol
is numerically valid, but are marked as extrapolative. Documented parameter
ranges are rates 0.1–1000, dwell time 0.001–3, μ 0.2–3.2, peak indentation
0–100, and total dimensionless time no greater than 300.

## Privacy and execution

Inference runs locally in the browser using TensorFlow.js with WebGL and CPU
fallback. The site has no backend, analytics, accounts, or input storage.

## Licensing

The model artifacts are licensed under CC BY 4.0; see `LICENSE-CONTENT`.