Distilling a Differentiable Plastic Potential from an Advanced Clay Model: Teacher–Student Knowledge Extraction in Bounding-Surface Plasticity

Advanced clay models capture anisotropy and stress history but are hard to calibrate and implement; compact bounding-surface models are practical but miss the anisotropic undrained response. This study distils knowledge from an advanced model into a compact one that remains a plasticity model. The teacher is MIT-S1 with published Boston Blue Clay parameters, whose own plastic strains are recorded; the student is the AJOP-T bounding-surface model with a fabric variable, written in a differentiable framework. Two closed forms are learned: a kinematic, fabric-dependent Nova-type plastic potential, identified by an inverse dilatancy analysis, and the exponent of the bounding plastic modulus. Splits, selection rule and pass criteria were fixed before fitting. Trained on 96 paths and selected by validation loss, the student reproduced all 32 held-out undrained paths with normalized error below 10%, predicted undrained strength within 10% for 94% of isotropic and 100% of K0 test paths, with a mean flow-direction mismatch of 7.8° against the teacher. Extension of K0 clay was not learned from compression data; two extension tests reduce the strength error on 17 unseen paths from about 47% to 11%. Trained only where the teacher matches laboratory data, a student predicted the unseen isotropic strength of Lower Cromer Till with a mean error of 12%, against 20% for the teacher. Trained on twelve synthetic MIT-S1 clays, it reproduced five laboratory clays about as well as the single-teacher student.

Authors

Publication Details

Journal
Preprints.org
Published
2026-11-17
DOI
https://doi.org/10.20944/preprints202611.1640.v1
Primary Topic
Geotechnical Engineering and Soil Mechanics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Distilling a Differentiable Plastic Potential from an Advanced Clay Model: Teacher–Student Knowledge Extraction in Bounding-Surface Plasticity

Sitthiphat Eua-apiwatch, Nopanom Kaewhanam, Sivarit Sultornsanee, Apichit Kampala et al.
Preprints.org
Geotechnical Engineering and Soil Mechanics
preprint

Distilling a Differentiable Plastic Potential from an Advanced Clay Model: Teacher–Student Knowledge Extraction in Bounding-Surface Plasticity

Sitthiphat Eua-apiwatch, Nopanom Kaewhanam, Sivarit Sultornsanee, Apichit Kampala, Thammanun Chatwong, Phongsakorn Wuttanachai, Phongsatorn Wuttanachai
preprint en

Abstract

Advanced clay models capture anisotropy and stress history but are hard to calibrate and implement; compact bounding-surface models are practical but miss the anisotropic undrained response. This study distils knowledge from an advanced model into a compact one that remains a plasticity model. The teacher is MIT-S1 with published Boston Blue Clay parameters, whose own plastic strains are recorded; the student is the AJOP-T bounding-surface model with a fabric variable, written in a differentiable framework. Two closed forms are learned: a kinematic, fabric-dependent Nova-type plastic potential, identified by an inverse dilatancy analysis, and the exponent of the bounding plastic modulus. Splits, selection rule and pass criteria were fixed before fitting. Trained on 96 paths and selected by validation loss, the student reproduced all 32 held-out undrained paths with normalized error below 10%, predicted undrained strength within 10% for 94% of isotropic and 100% of K0 test paths, with a mean flow-direction mismatch of 7.8° against the teacher. Extension of K0 clay was not learned from compression data; two extension tests reduce the strength error on 17 unseen paths from about 47% to 11%. Trained only where the teacher matches laboratory data, a student predicted the unseen isotropic strength of Lower Cromer Till with a mean error of 12%, against 20% for the teacher. Trained on twelve synthetic MIT-S1 clays, it reproduced five laboratory clays about as well as the single-teacher student.

Preprints.org
Quality education
Geotechnical Engineering and Soil Mechanics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon