Jianfeng Jin

Data preparation pipelines are a primary mechanism for improving data quality in machine learning workflows, transforming raw, error-prone tables into learning-ready data via sequential cleaning and feature transformation operators. However, automated pipeline construction remains computationally prohibitive due to the combinatorial complexity of operator sequences and the high cost of end-to-end evaluation. While Reinforcement Learning provides a principled discrete search paradigm, state-of-the-art (SOTA) Multi-DQN architectures suffer from three fundamental limitations: structural dissonance, where decoupled value estimators hinder long-horizon credit assignment; semantic detachment, where dataset context is treated as a superficial additive bias rather than strictly conditioning the agent's reasoning; and exploration inefficiency in a vast, sparse optimization landscape with many invalid pipeline states. To address these challenges, we propose FlowPipe , a unified framework that reformulates pipeline synthesis as conditional probabilistic flow generation over a directed acyclic graph. First, FlowPipe employs a Conditional Generative Flow Networks (C-GFlowNets) optimized via a Trajectory Balance objective, establishing a direct gradient path from terminal validation rewards to early actions to ensure holistic credit assignment. Second, to resolve semantic detachment, we introduce Deep Semantic Modulation via Feature-wise Linear Modulation (FiLM), which allows LLM-derived logical priors to multiplicatively modulate the policy's internal activation maps, thereby structurally adapting the decision logic to the dataset context. Finally, to mitigate exploration inefficiency, we incorporate failure awareness into the flow objective to prune semantically invalid states early and concentrate search mass on high-potential regions. Extensive experiments on two benchmark suites comprising 74 real-world datasets show that FlowPipe significantly outperforms SOTA baselines, improving accuracy by an average of 11.96% while achieving a 12.5× speedup in training convergence. The source code is available at https://github.com/KunyuNi/FlowPipe.

A small dataset of ∼300 datapoints of zinc (Zn) alloys were collected and 125 entries containing alloying elements, extrusion parameters (temperature (ET), speed (ES) and ratio (ER)), and mechanical properties (yield strength (YS), ultimate tensile strength (UTS), and final elongation (EL)) were selected. Machine learning models were applied to predict mechanical properties, in which random forest (RF) model exhibited the best performance and further validated by a new experimental sample of Zn‐0.05Mg‐0.5Mn, with the mean absolute percentage error (MAPE) less than 10%. [Correction added on 13 May 2025, after first online publication: In the preceding sentence, the value ‘12%’ has been changed to ‘10%’.]. Moreover, an empirical formula was induced by the clustering model (CL). Control over strain softening/hardening behavior was achieved through only process parameter adjustment. Finally, by combining multi‐objective genetic algorithm and RF models, the optimization alloy composition and extrusion parameters was carried out, targeting high‐strength, strength/plasticity synergy, and high plasticity for biodegradable purpose. A notable optimized scheme for strength/plasticity synergy in Zn‐0.20Mg‐0.60Mn (wt.%) achieves the YS of 303 MPa, UTS of 354 MPa, and EL of 25.1% with the MAPE less than 10%, and exhibits the strain‐hardening response, associated with ER of 16, ET of 170°C, and ES of 3.21 mm/s. [Correction added on 13 May 2025, after first online publication: In the preceding sentence, the value ‘345 MPa’ has been changed to ‘354 MPa’. and the value ‘3.33 mm/s’ has been changed to ‘3.21 mm/s’].

Bimodal grain structure (BGS) Mg alloys containing a high fraction of fine grains (FGs) and a low fraction of coarse grains (CGs) show a good combination of strength and plasticity. Here, taking the ZK60 alloy as an example, the influences of CG size, volume fraction, and texture intensity on mechanical properties and the hetero-deformation-induced (HDI) effect were examined using the Mori–Tanaka mean-field method combined with strain gradient theory of plasticity. The results indicate that the overall mechanical properties decrease with an increase in CG size because the limited HDI effect cannot compensate for the strength and plasticity decrease derived from larger CGs. A higher aspect ratio of CG along the loading direction can weaken the HDI effect and subsequently reduce the overall mechanical properties. Optimal comprehensive mechanical properties can be achieved when the CG volume fraction is approximately 30%. Furthermore, an increasing basal texture intensity in CG results in higher yield strength and lower ultimate tensile strength, while the uniform elongation reaches a maximum value when ~60% of CGs possess hard orientations with Euler angles of (0~30°, 0°, 0°).

The unique properties of graphene have attracted the interest of researchers from various fields, and the discovery of graphene has sparked a revolution in materials science, specifically in the field of two-dimensional materials. However, graphene synthesis’s costly and complex process significantly impairs researchers’ endeavors to explore its properties and structure experimentally. Molecular dynamics simulation is a well-established and useful tool for investigating graphene’s atomic structure and dynamic behavior at the nanoscale without requiring expensive and complex experiments. The accuracy of the molecular dynamics simulation depends on the potential functions. This work assesses the performance of various potential functions available for graphene in mechanical properties prediction. The following two cases are considered: pristine graphene and pre-cracked graphene. The most popular fifteen potentials have been assessed. Our results suggest that diverse potentials are suitable for various applications. REBO and Tersoff potentials are the best for simulating monolayer pristine graphene, and the MEAM and the AIREBO-m potentials are recommended for those with crack defects because of their respective utilization of the electron density and inclusion of the long-range interaction. We recommend the AIREBO-m potential for a general case of classical molecular dynamics study. This work might help to guide the selection of potentials for graphene simulations and the development of further advanced interatomic potentials.

The distinctive multi-ring structure and remarkable electrical characteristics of biphenylene render it a material of considerable interest, notably for its prospective utilization as an anode material in lithium-ion batteries. However, understanding the mechanical traits of biphenylene is essential for its application, particularly due to the volumetric fluctuations resulting from lithium ion insertion and extraction during charging and discharging cycles. In this regard, this study investigates the performance of pristine biphenylene and materials embedded with various types of hole defects under uniaxial tension utilizing molecular dynamics simulations. Specifically, from the stress‒strain curves, we obtained key mechanical properties, including toughness, strength, Young’s modulus and fracture strain. It was observed that various near-circular hole (including circular, square, hexagonal, and octagonal) defects result in remarkably similar properties. A more quantitative scaling analysis revealed that, in comparison with the exact shape of the defect, the area of the defect is more critical for determining the mechanical properties of biphenylene. Our finding might be beneficial to the defect engineering of two-dimensional materials.

The effect of the extrusion process on the microstructure, corrosion, and mechanical properties of Mg–Zn–Ca–Zr alloy has been investigated. Zn and Ca were both in a solid solution and only the Zr-rich phase was observed in the homogenized and extruded alloys. The Zr-rich phase was obviously refined after extrusion. The corrosion rate of the homogenized alloy decreased by about 25% after extrusion. This is because the refined Zr-rich phase was easier to cover with the deposited corrosion products, which reduced the cathodic reaction activity of the Zr-rich phase. The corrosion rate is similar for the alloys extruded at 320 °C and 350 °C since the size and distribution of the Zr-rich phase were not different in the two conditions. The alloy extruded at 320 °C has a smaller grain size and better comprehensive mechanical properties.

Beryllium has some unique properties and plays a key role in many special applications. However, Beryllium (α-Be) is of close-packed hexagonal (HCP) crystal structure, which has a strong anisotropic feature and limits its applications. In this work, diffusion behaviors of the self-interstitial atom (SIA) in α-Be at the temperature of 300–1100 K are studied using molecular dynamics simulations. It is observed that the diffusion mechanisms are not only dominated by the SIA jumps among the BO and BS sites on the basal plane, but also by the jumps among the C and O sites along the c-axis, which strongly depend on temperature. Diffusion behaviors of SIA can be divided into two stages with the temperature of 300–800 K and 800–1100 K, respectively, in which diffusion coefficient component of the c-axis (D _c) is higher than that of the basal plane (D _b) at first and then becomes closer to the D _b after 800 K, in consistent with diffusion mechanisms. When the temperature rises from 300 K to 1100 K, the total diffusion coefficient of SIA (D _t) increases gradually from 0.34 × 10⁻⁴ cm² s⁻¹ to 1.13 × 10⁻⁴ cm² s⁻¹. With the temperature increasing from 300 K to 1100 K, the anisotropy factor (η = D _c /D _b) of SIA diffusion drastically decreases from 1.76 to 1.01 in α-Be, while the η increases from 0.21 to 0.70 in α-Zr with the temperature from 500 K to 1100 K.

In this study, a small dataset of 370 datapoints of Mg alloys are selected for machine learning (ML), in which each datapoint includes five rare‐earth‐free alloying elements (Ca, Zn, Al, Mn and Sn), three extrusion parameters (extrusion speed, temperature and ratio), and three mechanical properties (yield strength [YS], ultimate tensile strength [UTS] and elongation [EL]). The ML algorithms, including support vector machine regression (SVR), artificial neural network, and other three methods, are employed, and the SVR has the best performance in predicting mechanical properties based on the components, and process parameters, with the mean absolute percentage error of YS, UTS, and EL being 6.34%, 4.19%, and 13.64% in the test set, respectively. The SVR model combined with multi‐objective genetic algorithm are successfully used to optimize mechanical properties of four extruded alloys from Mg‐Ca, Mg‐Ca‐Zn, Mg‐Ca‐Mn‐Sn, and Mg‐Ca‐Al‐Zn‐Mn series alloys, respectively, and the corresponding experimental results are in good agreement with the designed ones. Furthermore, new composition schemes are proposed from a wider range of elements and processing features to match the objectives of high‐strength, strength–ductility balanced, and high‐ductility Mg alloys, and the four‐, five‐ and six‐element alloying schemes are provided for the candidates of new‐generation wrought Mg alloys.

A combination of phase-field simulations and experimental validation is utilized to examine the effect of annealing tension on the microstructure evolution of 439 ferrite stainless steel (FSS). The study reveals the competing mechanisms of texture under tensile stress. Furthermore, a phase field model that incorporates anisotropic grain boundary (GB) energy and elastic energy is established. The microstructure of 439 FSS is created using a 3D reconstruction strategy based on the 2D electron backscatter diffraction characterization proposed in this work. Elastic constants are calibrated using actual alloy data and determined through molecular dynamics simulations. Finally, simulations of the grain coarsening process in 439 FSS are successfully achieved, considering both tensile stress and anisotropic GB energy effects. The results reveal that the presence of low-angle GBs deviates from Hillert model predictions in terms of grain size distribution and slows down the average grain size evolution over time. A significant deviation in the grain size distribution, compared to Hillert predictions, is observed in the textured system under tensile stress. The results of growth kinetics indicate that tensile stress promotes grain growth more than GB energy anisotropy retards microstructure evolution. Both experiment and simulation results consistently demonstrate that grains with //ND orientation experience a better growth proficiency compared to grains of other orientations under tensile stress. This investigation offers fresh insights into managing the ferritic microstructure of FSS to enhance its formability capabilities.