This study develops a dual global-local LSTM with multi-head cross-attention to forecast complete flexural curves of macro-basalt-fiber UHPC and ablates internal LSTM gates to test generalization to unseen mixes.
Key findings
- Several cell modifications improved error and training time. The no-forget-gate model reduced MSE by 57.7% and MAE by 30.7%, suggesting some gates may be unnecessary for the tested regime-changing mechanical response.
Why this matters globally
Forecasting full curves rather than peak strength can support toughness, cracking, and construction-material digital twins and may reduce destructive tests if externally validated.
Thai researcher contribution
A Thammasat University infrastructure research-unit author helped connect UHPC experiments with machine learning tailored to regime-changing material behavior.
Limitations to consider
Specimen and mix counts, split design, uncertainty, and alternative baselines are not stated. Leakage is possible if related curves cross splits, percentage improvement depends on baseline strength, and no physics constraint or blind external-lab test is reported.