Graded revision lets autonomous campaigns update symbolic laws of SiGeC semiconductors from mixed-fidelity evidence
Abstract
Autonomous materials campaigns combine published measurements, simulations at several levels of theory and curated compilations whose errors are known in kind but not in size. A value off the current law may be an error, a measurement under a special condition, or a lead to missing physics, and an outlier filter discards all three alike. Machine unlearning removes data from neural networks by costly or approximate retraining, and provenance records locate data origins without identifying which conclusions to revise. We tested whether nine graded revision actions on sourcetagged evidence, from labelling or reweighting rows to forgetting them, confine each correction to conclusions named in advance. The actions operate on symbolic laws fitted by HG-SR, a symbolic-regression engine whose terms record their supporting evidence, and were applied to the lattice thermal conductivity of Si 1-x Ge x and SiGeC, which limits heat removal from SiGe transistors. The pipeline corrected an undesigned unit error, caught mislabelled and amorphous simulation rows from their structure, split the literature by microstructure and issued a conductivity law with a declared validity domain. Of twelve decisions on flagged values, four found errors; the other eight kept the value, five as special conditions (microstructure, another alloy, melting, box size, nonlinear response) and three as leads to physics the law lacked. For SiGeC it showed that carbon leaves the lattice as SiC and named in advance the one weak input of the germanium-saving claim; pre-registered tests of that input changed only the named conclusions for a leg operated from 300 to 1100 K, leaving its composition and carbon content unchanged. Because the SiC scattering contrast fell with particle radius roughly as R −4 , 2.5 nm SiC replaces 14–16% of the germanium at equal lattice thermal conductivity (8–9% at 5 nm). Corrections become targeted errata, not rebuilds, and less germanium eases demand for a by-product metal of constrained supply.
Related
- Covering and loss-tolerant batch designs guarantee what autonomous campaigns sample in SiGeC semiconductors
- Hash-committed pre-registration and exact unlearning make belief revisions in autonomous SiGeC semiconductor campaigns checkable
- CampChem: Rubric-Grounded Adaptive Campaigns for Multi-Round Organic Synthesis with Residual Spectral Observation
- Agentic Workflow for High-Throughput Synthesis Campaigns: Mapping the Persistent Phosphor Ca2ZnSi2O7:Eu2+,Nd3+
- Same Raw Data, Different Numbers: Locked Analysis Rules Make Film Descriptors of Organic Semiconductors Reproducible
BibTeX
@misc{ackolf2026graded,
title = {{Graded revision lets autonomous campaigns update symbolic laws of SiGeC semiconductors from mixed-fidelity evidence}},
author = {Kinston Ackölf and Taylor D. Sparks},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010672/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010672/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Graded revision lets autonomous campaigns update symbolic laws of SiGeC semiconductors from mixed-fidelity evidence
AU - Kinston Ackölf
AU - Taylor D. Sparks
PY - 2026
DA - 2026/10/19
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010672/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010672/v1
AB - Autonomous materials campaigns combine published measurements, simulations at several levels of theory and curated compilations whose errors are known in kind but not in size. A value off the current law may be an error, a measurement under a special condition, or a lead to missing physics, and an outlier filter discards all three alike. Machine unlearning removes data from neural networks by costly or approximate retraining, and provenance records locate data origins without identifying which conclusions to revise. We tested whether nine graded revision actions on sourcetagged evidence, from labelling or reweighting rows to forgetting them, confine each correction to conclusions named in advance. The actions operate on symbolic laws fitted by HG-SR, a symbolic-regression engine whose terms record their supporting evidence, and were applied to the lattice thermal conductivity of Si 1-x Ge x and SiGeC, which limits heat removal from SiGe transistors. The pipeline corrected an undesigned unit error, caught mislabelled and amorphous simulation rows from their structure, split the literature by microstructure and issued a conductivity law with a declared validity domain. Of twelve decisions on flagged values, four found errors; the other eight kept the value, five as special conditions (microstructure, another alloy, melting, box size, nonlinear response) and three as leads to physics the law lacked. For SiGeC it showed that carbon leaves the lattice as SiC and named in advance the one weak input of the germanium-saving claim; pre-registered tests of that input changed only the named conclusions for a leg operated from 300 to 1100 K, leaving its composition and carbon content unchanged. Because the SiC scattering contrast fell with particle radius roughly as R −4 , 2.5 nm SiC replaces 14–16% of the germanium at equal lattice thermal conductivity (8–9% at 5 nm). Corrections become targeted errata, not rebuilds, and less germanium eases demand for a by-product metal of constrained supply.
ER -