Carbon segregation is a persistent defect in continuous casting of large-diameter steel billets, leading to deteriorated mechanical properties and compromised service reliability. Conventional empirical or machine learning models generally estimate segregation indices but cannot resolve local variations of carbon distribution across billet sections. In this work, a microstructure-informed convolutional neural network (CNN) framework is proposed to predict and map carbon segregation in 600 mm round 42CrMo steel billets. A comprehensive dataset comprising microstructural images and corresponding carbon content measurements was established. The customized CNN achieved a testing accuracy of 81.3% with a mean absolute error of 0.012 wt.% and showed good robustness in out-of-sample validation. Compared with transfer learning models (VGG16, VGG19, etc.), the customized architecture exhibited superior generalization on this domain-specific dataset. Contrast-enhanced imaging significantly improved predictive performance, while Gradient-weighted Class Activation Mapping visualizations highlighted key microstructural regions correlated with carbon distribution, providing interpretability. This study demonstrates a proof-of-concept methodology to achieve quantitative mapping of segregation patterns in large-diameter 42CrMo billets, offering a complementary tool to traditional metallurgical analysis and providing a workflow that may support future data-driven research on segregation formation mechanisms and process optimization in steel casting when extended to additional steels and casting conditions.

This work presents a novel two-step composite femtosecond laser drilling process integrating spiral punching and spiral finishing for high-precision micro-hole fabrication in 304 stainless steel fuel injector nozzles, aiming to resolve the conflict between machining accuracy and thermal defects caused by the lack of optimized process parameters and routes. Based on the two-temperature model (TTM), the critical physical parameters governing processing quality in both percussion drilling and spiral drilling are quantitatively determined, including single-pulse energy, laser beam waist, scanning speed, and defocusing amount. Experimental results demonstrate that single-mode drilling cannot meet the required quality and accuracy due to unavoidable thermal defects and geometric distortion. In contrast, the proposed composite process first forms a functional channel via spiral punching, then removes the recast layer and precisely refines the hole diameter through spiral finishing. Compared with the conventional one-step spiral drilling, this composite strategy achieves remarkable quantitative improvements: the K-factor taper of fuel injection holes is converted from a positive cone to a negative cone, the inner wall surface roughness is reduced by 31%, and the phase transformation around holes is decreased by 62%. A quantitative process mapping relationship between percussion drilling and spiral drilling is further established, where the material removal volume per pulse shows strong consistency between 25-pulse percussion craters and 5-pulse linear grooving craters. This relationship provides reliable theoretical support for the quantitative optimization of the composite drilling process. The proposed two-step femtosecond laser drilling method provides an effective and practical solution for high-accuracy micro-hole manufacturing.

The use of flexible compositional ranges is emerging as a recycling-oriented alloy design strategy that accommodates mixed-scrap variability while maintaining targeted performance. Here, we experimentally validate a compositionally flexible, scrap-tolerant austenitic stainless steel design using controlled laboratory remelting of realistic stainless-steel scrap mixtures, where mixed scraps introduce both principal element scatter (Cr, Ni, Mn) and coupled impurity inputs. Using direct remelting of standard (316L, 201) scraps and non-standard scraps enriched in C, N, and Nb, we quantify how impurity-driven compositional variability controls phase constitution, strength stability, and pitting corrosion within a defined flexible composition space. Alloys produced from mixed 316L and 201 scraps remained fully austenitic and mechanically stable despite up to 4 wt% variation in principal elements, achieving a yield strength of 220 ± 20 MPa. Their pitting resistance increased with scrap-inherited Mo, giving pitting potentials of 220–320 mV (vs. SCE). Mass balance analysis further shows that C is largely retained during laboratory remelting, whereas N retention is strongly dependent on the scrap route: standard-scrap alloys show substantial N loss, while miscellaneous-scrap-containing alloys exhibit higher apparent N retention associated with Nb, W and V-bearing inputs. After considering N partitioning into Nb-rich M(C, N)-type precipitates, the strengthening analysis indicates that solid-solution strengthening remains the dominant strengthening contribution. These results demonstrate that compositionally flexible alloy design should be extended from a nominal Cr-Ni-Mn window to a scrap-aware framework that couples principal-element tolerance with secondary-element retention, matrix/precipitate partitioning and impurity consistency.

The design of tough alloys as maraging steels was mainly explored by integrated computational materials engineering and domain-knowledge guided artificial intelligence (AI), relying critically on the introduction of accurate physical mechanisms. However, the discovery and quantification of mechanisms remain dependent on time-consuming experimental characterization and expert interpretation, which are prone to subjective bias. These challenges are limiting the accuracy and efficiency of the design of alloys with satisfactory toughness. Here, we propose a causal reinforcement learning framework based on dynamic knowledge graphs for the autonomous discovery of mechanisms directly from reported literature data. In this framework, a reinforcement learning agent iteratively prunes the graph, guided by causal principles to identify influential relationships, while predictive improvement is used as the reward. Applied to ultrahigh-strength maraging steels, the framework learned from high-quality literature data and autonomously uncovered the mechanistic chain “Mo content → interfacial energy → toughness”, which is easily overlooked in conventional approached. Guided by this mechanism, the AI framework identified multiple alloy candidates with excellent strength–toughness synergy. Moreover, the model showed strong generalization ability, maintaining high predictive accuracy in both out-of-domain and data-scarce regimes where conventional models suffered severe performance degradation. By this knowledge-led AI strategy, the toughness-controlling mechanisms can be automatically summarized from literature data, which can greatly reduce the complexity of revealing mechanisms through experimental methods and accelerate the development efficiency of tough alloys. This method can be extended to other performance designs involving controversial mechanisms.

The transient temperature field in laser powder bed fusion (LPBF), governed by complex interactions among multiple process parameters, critically determines part quality and properties. Conventional numerical methods are often hindered by limited computational efficiency and poor generalizability across parameter spaces and material systems, impeding multi-parameter co-optimization. To overcome these challenges, this study proposes a parameterized physics-informed neural networks (PINNs) framework that integrates process parameters and material thermal properties as inputs, establishing a unified model for temperature-field prediction across diverse alloy systems. Validated on IN718 alloy, the model achieves high-fidelity prediction of melt-pool morphology, with width and depth prediction accuracy exceeding 90% and 88%, respectively, showing close agreement with experimental and simulation benchmarks. Transferred to the AMSC-1Nb system through material parameter substitution alone, the model maintains robust performance, delivering depth and width predictions with mean accuracy above 92% and 90%. By enabling accurate prediction across a continuous parameter space and demonstrating cross-material transferability, the proposed framework extends the practical applicability of PINNs in LPBF and offers an efficient, physics-informed tool for digital process optimization and intelligent manufacturing.

Microstructural images often contain complex backgrounds and subtle target features, which can make classification models vulnerable to interference from irrelevant information. To address this issue, this study proposes a target-region-prior-guided framework for microstructural image analysis and evaluates it using steel inclusion classification as a practical application task. First, a pretrained Segment Anything Model was used to generate candidate inclusion regions, followed by an image-relative brightness-deviation refinement strategy based on a shared threshold to construct inclusion masks. The refined masks were then fused with the original images by pixel-wise multiplication, thereby attenuating much of the background content while preserving the grayscale and morphological information of the target regions. Based on this fused input, a Mask-Guided Fusion InceptionV3 (MGF-InceptionV3) model was developed for inclusion classification. Within the present dataset, the proposed method achieved an average classification accuracy of 97.4 ± 2.4% and a macro-averaged F1 score of 0.975. Comparative experiments across six backbone networks further showed that the fusion-based strategy generally performed better than input modes based only on the original image or only on the mask. In addition, Grad-CAM++ visualizations provided qualitative support that the fused input was associated with more target-centered responses in the inclusion regions. Taken together, these results suggest that incorporating target-region priors into input construction is a useful strategy for reducing background interference and improving inclusion classification within the present dataset.

For sustainable alloy design, unified-composition approaches offer an effective route to deliver multiple performance levels while reducing chemistry complexity. Quenching and partitioning (Q&P) steels are widely used advanced high-strength steels, yet their grade development typically relies on distinct, grade-specific chemistries, complicating welding and hindering efficient recycling. This study proposes a physics-informed machine learning framework for unified-composition Q&P steel design, enabling multiple strength grades from a single alloy via heat-treatment tuning. A physics-guided property-bridging model integrates metallurgical descriptors with near-high-throughput hardness data, transferring knowledge from hardness to tensile properties under sparse tensile labels. This approach enables improved prediction of tensile strength and elongation from limited tensile datasets. A multi-objective genetic algorithm then explores the composition-process space to identify one alloy that meets the Q&P980, Q&P1180, and Q&P1380 grade targets via processing adjustments. Compared to a purely data-driven baseline, the framework substantially improves tensile prediction accuracy (R2 up to 84% vs. 74%) while maintaining better model stability under data-sparse conditions. An experimental validation demonstrates that the same chemistry can achieve ∼980, ∼1180, and ∼1380 MPa tensile strengths with suitable ductility under different Q&P schedules. Overall, the physics-informed framework exemplifies a paradigm for sustainable alloy design that reduces chemistry variants while simplifying recycling.

Metal additive manufacturing (MAM) offers significant advantages in near-net shape production of complex parts. However, the intricate relationship between process parameters, microstructure, and mechanical properties makes quality control challenging. Machine learning (ML) offers a powerful framework to address this issue. Data scarcity remains the primary bottleneck that restricts the application of ML in MAM. We have conducted a comprehensive exploration of the issues and solutions related to data scarcity in MAM and discussed various techniques for data generating, sampling, and fusion. These range from high-throughput experiments and simulations to advanced methods such as active learning and multi-fidelity fusion technology. In addition, we address the interpretability challenge in MAM by highlighting the trade-off between data, accuracy, and physical consistency. By classifying methods of domain knowledge integration, we provide a systematic guide to their advantages and constraints in enhancing model reliability. Finally, a roadmap for ML in MAM is proposed, highlighting the synergy between knowledge graph-driven RAG agents, high-fidelity digital twins, and embodied intelligence. This path aims to achieve autonomous manufacturing through self-improving closed-loop systems, accelerating the transition to fully intelligent, data-driven production.

Selective mid-infrared radiative materials hold significant potential for applications in infrared stealth, radiative cooling, and energy utilization. However, the vast and complex design space presents a considerable challenge in meeting the intricate spectral requirements for infrared stealth compatible with thermal management using multilayer thin-film metamaterials. To address this challenge, this study combines artificial intelligence with finite element simulation to enable high-throughput screening of structural parameters, successfully designing a multifunctional thin-film metamaterial based on a four-layer Ge-Ti structure. The thin-film metamaterial achieves low emissivity in the infrared atmospheric window while exhibiting high emissivity in the non-atmospheric window (e5-8 & micro;m = 0.872) without introducing additional structural complexity. Through AI-driven structural optimization, the material further enables broadband radiative thermal management across the 5-14 & micro;m range. The flexible thin-film metamaterial is characterized by a straightforward fabrication process, making it highly suitable for scalable and cost-effective mass production. Combined with its exceptional tunability of mid-infrared radiative properties, it holds great promise for applications in thermal management, energy utilization, and military infrared stealth. Additionally, the proposed framework establishes a scalable and transferable AI-assisted design paradigm that enables materials researchers without specialized expertise in artificial intelligence to efficiently design thin-film systems with application-specific radiative properties.

Reliable AI-driven materials design depends on benchmark datasets that connect chemical composition, processing routes and properties. In steel research, such data remain scattered across heterogeneous literature, limiting reproducible model development and fair comparison. Here we present SteelProBench, a large-scale benchmark for steel materials informatics that provides 14,825 DOI-traceable, machine-readable material-state records linking material designation, chemical composition, processing routes and mechanical properties. The records cover diverse steel families and preserve physically meaningful strength-ductility and processing-pattern distributions. To construct this resource at scale, we develop a reproducible collaborative multi-large language model (LLM) extraction pipeline that combines metallurgy-aware validation, reliability-weighted fusion and post-extraction normalization. Expert audit shows high record-level fidelity, with an overall accuracy of 98.38%. Validation against an independent reference set shows strong agreement with expert records and identifies additional valid records missed by manual curation. We further demonstrate utility by fine-tuning open-source LLMs to infer mechanical properties from composition and processing descriptions. SteelProBench reconciles dataset scale with data fidelity and provides a traceable benchmark for AI-assisted steel design.

This study systematically investigates the influence of residual stress on the microbiologically influenced corrosion (MIC) and mechanical properties of M54 martensitic steel in the presence of Desulfovibrio vulgaris. Controlled residual stress levels were applied, including low stress (LS,-215 f 10 MPa), moderate stress (MS,-387 f 19 MPa), and high stress (HS,-468 f 23 MPa). The results indicated that residual stress significantly aggravated MIC susceptibility. In the biotic medium, the LS coupon exhibited superior MIC resistance, with a corrosion current density of 6.2 f 0.2 mu A cm-2, weight loss of 1.1 f 0.12 mg cm-2, and average pit depth of 3.6 f 0.3 mu m. In contrast, the HS coupon showed severe corrosion damage, with corresponding values of 19.2 f 2.1 mu A cm-2, 2.5 f 0.12 mg cm-2, and 6.1 f 0.5 mu m. Corrosion preferentially initiated at prior austenite grain boundaries (PAGBs), which served as electron donation sites for D. vulgaris. Stress concentration zones in the HS coupon further promoted intergranular attack and cracking. More critically, the synergistic effect of residual stress and microbial activity resulted in severe mechanical degradation, with the HS coupon suffering reductions of 18.9 % in yield strength, 10.1 % in tensile strength, and 27.3 % in elongation compared to the LS coupon in abiotic medium. These findings underscore the detrimental role of residual stress in accelerating MIC and mechanical failure, emphasizing the importance of stress management for enhancing the durability of martensitic steels in microbial environments.

Balancing the transformation-induced plasticity (TRIP) strengthening against the inherent brittleness of deformation-induced martensite remains a key challenge in achieving optimal strength-toughness synergy in austenitic steels. In this study, the effect of the temperature-dependent martensitic transformation on the impact toughness of an Fe-Cr-Ni austenitic steel was studied. With increasing temperatures, the toughening mechanisms span from detrimental TRIP effect to sustained twinning-induced plasticity (TWIP)-assisted TRIP effect to single TWIP effect. The peak toughness is attributed to the unique impact fracture mechanism of the concurrent of TWIP and TRIP effects, which collectively relaxes stress concentration and enhances strain hardening capacity.

Machine learning offers a promising approach to design high-performance alloys for laser additive manufacturing, by bypassing convoluted physical models and identifying correlations among composition, processing, microcracks/porosity and properties. However, conventional machine learning methods face limitations, e.g., overfitting or unreasonable design results, due to reliance on large, high-quality datasets. Here, we introduce a generic framework operable with smaller experimental datasets, by integrating knowledge-informed graph modeling alongside data uncertainty quantification. The generic physical-metallurgy knowledge and the stochasticity of experimental defect distributions from produced material are rationally balanced. To validate the approach, we detail the development of a new defect-free Ni superalloy possessing excellent laser printability, thermal stability, and high mechanical strength. Mechanism mining revealed a possible origin for this performance, which was confirmed by atom probe tomography. Subsequently, we developed a new laser-printable aluminum alloy through the same approach, highlighting the framework’s potential to accelerate next-generation alloy design for additive manufacturing. The authors report a machine-learning framework that combines a physical-metallurgy knowledge graph with uncertainty analysis to design alloys for laser powder bed fusion from small datasets, yielding crack-free nickel and aluminum alloys with high performance.

Carbide-free bainite (CFB) steel, consisting of bainitic laths and metastable retained austenite (RA), offers excellent static properties and fatigue resistance, making it a potential replacement for conventional axle steels. However, the relationships among its microstructure, fatigue life under various stresses, and fatigue crack growth (FCG) behavior remain unclear. This study designed two CFB axle steels, AT350 and AT300, by austempering at 350 degrees C or 300 degrees C followed by low-temperature tempering, producing different bainitic lath sizes and RA stabilities. Their mechanical properties, high-cycle fatigue (HCF) performance, and FCG behavior were systematically investigated. The AT300 sample featured finer bainite laths (average thickness: 231 nm) and higher RA stability than that of the AT350 sample, due to prior austenite grain segmentation by primary martensite. Consequently, AT300 achieved higher tensile strength (1575 MPa) and HCF strength (760 MPa, fatigue limit at 107 cycles). Under low-stress fatigue, the AT300 sample exhibited longer life and slower crack growth than the AT350 sample. Under high-stress fatigue, however, the fatigue lives of the AT350 and AT300 samples were nearly equivalent, as AT350 ' s greater ductility and secondary cracking compensated for its lower strength. In the FCG regime, AT300 showed a wider resistance plateau and a lower Paris exponent, indicating superior crack growth resistance. This enhancement arises from fine bainitic laths deflecting cracks and stable RA reducing martensitic transformation at the crack tip, thereby avoiding brittle martensite channels and absorbing strain energy. These findings provide valuable theoretical and experimental guidance for designing axle steels with superior fatigue resistance.

Bainite is a crucial microstructure in steel, and its transformation kinetics plays a key role in microstructural control and property optimisation. Conventionally, bainite formation initiates at austenite grain boundaries and continues through autocatalytic nucleation at the newly formed bainitic ferrite/austenite interfaces. However, recent studies have revealed that the presence of martensite prior to the austenite-to-bainite transformation can significantly alter the transformation behaviour. In this work, within the framework of the displacive mechanism, a new isothermal kinetic model is proposed by introducing the catalytic effect of martensite/austenite interfaces to characterise the influence of pre-existing martensite on bainite transformation in high-Si steels. The model was validated through dilatometry experiments performed on two Fe-C-Mn-Si steels with different Mn contents, under both above-Ms and below-Ms temperature conditions. The results demonstrate that bainite transformation exhibits an incubation period above Ms, whereas below Ms, the pre-existing martensite markedly shortens the incubation stage and accelerates the initial transformation rate. There is a high level of agreement between the model predictions and the experimental observations, successfully capturing the key features of the transformation behaviour, including transformation rate, incubation period, final fraction, and the incomplete transformation phenomenon.

This study investigates how precipitation evolution influences the microbiologically influenced corrosion (MIC) resistance and mechanical properties of M54 maraging steel, with emphasis on post-corrosion performance. Four aging conditions were examined: precipitation-free with partial (AT2-1h, 200 degrees C-1h) and near-complete (AT5-1h, 520 degrees C-1h) stress relief, under-aged with moderate precipitation (AT5-5h, 520 degrees C-5h), and peak-aged with abundant nano-carbides (AT5-10h, 520 degrees C-10h). In the abiotic medium, all coupons showed comparable corrosion resistance. However, under Desulfovibrio vulgaris exposure, short-term aging enhanced MIC resistance via stress relief (AT5-1h optimal), while prolonged aging introduced carbide-induced micro-galvanic corrosion, deteriorating MIC resistance (AT5-10h worst). Mechanically, the AT5-10h coupon exhibited the highest intrinsic strength due to carbide precipitation strengthening, but experienced the most severe performance loss owing to its inferior MIC resistance, with its elongation dramatically dropping from 10.3% to 4.7% after MIC exposure. In contrast, benefiting from better MIC resistance, the AT5-5h coupon retained a yield strength of 1678 f 21 MPa, a tensile strength of 1847 f 24 MPa, and an elongation of 10.4 f 0.2% after MIC exposure. Consequently, in SRB-containing environments, seeking the optimal balance between mechanical properties and MIC resistance, embodied by the AT5-5 h coupon, is the key to achieving extended service life for maraging steel.

The use of domain knowledge to guide artificial intelligence has been widely applied in materials science. However, in extreme corrosion environments controlled by multiple factors, effectively integrating appropriate domain knowledge into machine learning models has become a key challenge. Taking the sulfide stress corrosion cracking (SSCC) behavior of oil country tubular goods in oil and gas well environments as an example, this study developed a mechanism identification and prediction model based on a graph convolutional network (GCN). The model constructs an initial database incorporating composition and processing parameters, as well as various mechanistic information obtained from density functional theory (DFT) calculations, including grain boundary separation work, Young's modulus, and Vickers hardness other mechanistic information. On a small dataset containing only 141 samples, the prediction accuracy of the GCN model improved from 0.83 to 0.873 after introducing a knowledge graph. Subsequently, an algorithmically constrained scoring method was employed to identify the dominant influencing factors of SSCC and to construct a more rational mechanism graph. By identifying the dominant mechanism and updating the graph, the prediction accuracy of the model was further improved to 0.905, and grain boundaries were successfully identified as the more dominant mechanism for the present dataset. This approach may provide a feasible pathway for predicting complex properties and identifying key physical factors among multiple candidate mechanisms in materials science.

The trade-off between printability and mechanical properties is a major challenge in additive manufacturing (AM) of high-performance aluminum (Al) alloys. Here we propose a new machine learning-driven alloy design framework integrating graph attention networks with a physical-metallurgy knowledge graph, which synergistically incorporates complex factors including the mechanistic information related to processing parameters and thermodynamic information related to alloy compositions. By further incorporating a genetic algorithm, a novel Al alloy KG-AMAA has been developed as the optimization of relative density, as confirmed by the robustness of printing parameters in fabricating crack-free samples. The new alloy KG-AMAA exhibits superior strength and ductility compared with most existing AM-ed Al alloys. Microstructural characterization revealed that the coexistence of multiple precipitate systems in the designed alloy, such as the primary precipitate Al3Sc, and the eutectic phases AlSc2Si2 and Si, which collectively contributed to the enhanced printability, together with Al2Cu strengthening phases, established exceptional strength-ductility synergy. This novel machine learning paradigm, integrating physical metallurgy knowledge, offers innovative approaches and great potential for alloy design for additive manufacturing.

Surfaces with gradient wetting properties are crucial for alleviating water scarcity and efficient recycling, as it can be used for directed transport and collection of water. However, a major challenge is how to efficiently manufacture multi-scale micro-nano structures to increase surface wetting gradients. In this study, a composite processing that combines nanosecond laser oblique incidence and thermal oxidation to fabricate microgroove-nanowire (CuO) hierarchical structures with gradient wettability is proposed. Laser oblique incidence scanning is employed to fabricate microgroove structures with gradient geometric dimensions and chemical composition on the Cu surface, which in turn induced gradient wetting properties and facilitated the directed movement of droplets. The effect of scanning times on gradient structure and its wetting properties is also discussed in detail. To further enhance the directional flow distance of droplets, dense CuO nanowires are grown on the surface of microgrooves through thermal oxidation treatment, forming a micro-nano dual scale structure. The growth mechanism of nanowires is revealed, and the effects of thermal oxidation temperature and duration on nanowire growth and gradient wetting properties are discussed in detail. The gradient in contact angles, in conjunction with the variation in energy barriers in different directions, leads to more pronounced anisotropic wetting. Compared to the single microgroove structures, the directional flow distance of microgroove-nanowire dual scale structure is increased by about 50

Stacking fault energy (SFE) significantly influences plastic deformation, strength, and processing performance, making accurate assessment and prediction of SFE essential for material design and optimization. Traditional SFE calculations mainly rely on experimental measurements and thermodynamic theories, with the former usually being time-consuming and the latter limited in applicability at different compositions. To overcome these limitations, this study proposes a machine learning (ML) strategy introducing physical metallurgy (PM) parameters relevant to SFE, aiming to achieve robust predictions. Specifically, this study evaluates three methods for introducing PM information into ML (as an input, an intermediate parameter, and a transfer source), with transfer learning as the best strategy. Initially, various PM parameters were calculated based on alloy composition and temperature, and subsequently used as inputs to train a convolutional neural network (CNN). This source model was then transferred to the SFE prediction model. The results from the model transfer using different PM information show that incorporating phase-transformation driving force (DF) as a source model for SFE prediction provided the most accurate and reliable results. This approach of introducing PM parameters into ML significantly improves the predictive capability of SFE models, offering a new perspective and solution for the prediction of SFE. Furthermore, this method may also be applicable to the prediction of other material properties during material design and optimization.