Testing predictive-state sufficiency and recursive closure in additive-manufacturing microstructure models
Abstract Most additive-manufacturing models predict final microstructure from process parameters or thermal-history descriptors. We ask a different question: how much of a material’s thermal history must be carried forward to predict what the next heat cycle will do to it? We call this compact, continuously updated information a predictive material state. In Ti–6Al–4V simulations we hid the phase fractions and inferred state coordinates from noisy total-alpha measurements. Two coordinates reduced identification RMSE by 27.1% relative to one; for held-out continuation prediction, one coordinate lay within the prespecified 0.001 RMSE margin of the best tested dimension. Transferred to a separately formulated kinetic model with only the readout recalibrated, a three-coordinate state cut median prediction error by 87.1% relative to ten conventional thermal-history descriptors with access to the entire past. Carried recursively through six unseen thermal segments with no intervening measurement, a two-coordinate state stayed 86.8% below the same descriptor baseline. A recurrent network forecast terminal phase fractions; its hidden state was not evaluated for predictive sufficiency or recursive closure. These computational results support task-dependent predictive compression within the tested input domain and tolerance.
Authors
- В. И. Копылов (ORCID: https://orcid.org/0000-0002-1789-3226)
- Anatoliy V. Zavdoveev (ORCID: https://orcid.org/0000-0003-2811-0765)
- Ji Junwen (ORCID: https://orcid.org/0000-0001-7959-753X)
- Jie Liu
Institutions
- E.O. Paton Electric Welding Institute (UA)
- National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute” (UA)
- Nanjing University of Industry Technology (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-16
- DOI
- https://doi.org/10.1038/s41598-026-77131-2
- Primary Topic
- Machine Learning in Materials Science
- Type
- article
- Field-Weighted Citation Impact
- 0.00