Scholarlink

<p>Abnormal grain growth (AGG) in powder-processed Ni-base superalloys arises from coupled, multiscale interactions during isothermal forging (ITF) and super-solvus heat treatment (HT), eluding reliable prediction. We present a high-dimensional phenomenological metallurgy framework that couples domain knowledge-informed feature engineering with stage-wise neural networks to learn bidirectional process–structure mappings directly from EBSD/EDS datasets. Automated computer vision (CV) pipelines yield 71 physically interpretable descriptors for two-phase <i>γ</i>/<i>γ</i>′ microstructures after forging and 21 descriptors for single-phase <i>γ</i> microstructures after HT. Two combined neural networks (NNs) map strain to an intermediate (after ITF) and subsequently to a final microstructure state (after HT). Physically constrained descriptor-space augmentation expands feature coverage while preserving metallurgical realism and enables stable training without exposing NNs to raw ground-truth data. The composed model evaluates dense parameter grids at microsecond throughput per point and reproduces experimental trends with MAPE ≈ 2 to 15&#xa0;pct across descriptors. The methodology is capable to recover process parameters directly from experimentally observed microstructures and generalizes to an experimental test dataset with distinct thermomechanical history. Process maps expose a critical window at low to intermediate strains (≈0.6) combined with slow heating ramps, where the susceptibility for AGG peaks, while rapid heating suppresses AGG across broad strains. Mechanistically, the largest <i>γ</i> grains in <i>γ</i>′ voids accumulate significant stored plastic strain energy, seeding selective growth once <i>γ</i>′ pinning collapses during HT. Interpretable descriptors, combined with hierarchical NN composition, thus convert microscopy-anchored measurements into predictive and diagnostic process maps, revealing the causal pathway to AGG and providing actionable levers for AGG-resistant manufacturing of turbine-disk alloys.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Neural Network Modeling of Process–Structure Relationships and Abnormal Grain Growth in Polycrystalline Ni-Base Superalloys Through Domain Knowledge-Informed Feature Engineering

  • 期刊论文

  • 发表日期：

    2026年2月10日

查看全文
  • Pascal Thome,
  • Luis F. Arciniaga,
  • Cécile F. Hohenadel,
  • Owen Lowery,
  • Sammy Tin

摘要

Abnormal grain growth (AGG) in powder-processed Ni-base superalloys arises from coupled, multiscale interactions during isothermal forging (ITF) and super-solvus heat treatment (HT), eluding reliable prediction. We present a high-dimensional phenomenological metallurgy framework that couples domain knowledge-informed feature engineering with stage-wise neural networks to learn bidirectional process–structure mappings directly from EBSD/EDS datasets. Automated computer vision (CV) pipelines yield 71 physically interpretable descriptors for two-phase γ/γ′ microstructures after forging and 21 descriptors for single-phase γ microstructures after HT. Two combined neural networks (NNs) map strain to an intermediate (after ITF) and subsequently to a final microstructure state (after HT). Physically constrained descriptor-space augmentation expands feature coverage while preserving metallurgical realism and enables stable training without exposing NNs to raw ground-truth data. The composed model evaluates dense parameter grids at microsecond throughput per point and reproduces experimental trends with MAPE ≈ 2 to 15 pct across descriptors. The methodology is capable to recover process parameters directly from experimentally observed microstructures and generalizes to an experimental test dataset with distinct thermomechanical history. Process maps expose a critical window at low to intermediate strains (≈0.6) combined with slow heating ramps, where the susceptibility for AGG peaks, while rapid heating suppresses AGG across broad strains. Mechanistically, the largest γ grains in γ′ voids accumulate significant stored plastic strain energy, seeding selective growth once γ′ pinning collapses during HT. Interpretable descriptors, combined with hierarchical NN composition, thus convert microscopy-anchored measurements into predictive and diagnostic process maps, revealing the causal pathway to AGG and providing actionable levers for AGG-resistant manufacturing of turbine-disk alloys.