Accurate early-stage prediction of molecular properties and toxicity is critical for reducing cost and attrition in drug discovery. Here, we develop and evaluate an optimized 3D-MoRSE (OPT3D) molecular representation that introduces a tunable distance-scale factor to better capture informative interatomic distance regimes. With simple traditional machine-learning models, for example, random forest (RF), OPT3D achieves an average (10 times) test root mean squared error (RMSE) of 0.942 (ESOL), 0.940 (Lipophilicity), and 2.108 (FreeSolv), and AUC of 0.824 (BACE), 0.878 (BBBP), 0.621 (SIDER), 0.828 (Tox21), and 0.717 (ToxCast). These results are competitive with recent state-of-the-art predictions that rely on more complex architectures. The prediction performance of this set of descriptors can be further improved by more advanced models: stacked ensembling further reduces regression errors and maintains strong classification performance, and the adaptive checkpointing with specialization neural model can also improve performance relative to simple traditional machine-learning models. Performed perturbation-based tests also show that the OPT3D descriptor exhibits strong robustness under moderate noise. Our results highlight that developing high-quality molecular representations is as important as model innovation, which has been intensively pursued but has yielded limited performance gains.
Abstract In multiscale modeling of ferroelectric materials, a fundamental challenge is to transfer discrete atomistic information into a continuum phase-field model (PFM) while retaining an accurate description of mesoscale behavior. In this paper, a physics-informed neural network (PINN) framework is developed, in which the loss function consists of MD-data penalty and partial difference equation (PDE) constraints. By minimizing the loss function, the model not only reconstructs the polarization field along with the full coupled electromechanical response, including strain, stress, electric field, and energy landscape at the continuum scale, but also identifies critical parameters required for the PFM, including the characteristic energy density, characteristic length factor, anisotropy factor, and Landau polynomial coefficients. By using the PINN-predicted parameters to solve the corresponding PDEs within a finite element framework, we perform a cross-verification showing that the two numerical implementations yield consistent results. In addition, the transferability of the learned parameters is further evaluated through extended three-dimensional PF simulations under external tensile and bending loading, in which the domain expansion, shrinkage, annihilation, and fragmentation are in agreement with the corresponding experimental observations. This framework provides an effective methodology for multiscale bridging between atomistic and continuum descriptions for perovskite ferroelectric materials.
Rare earth permanent magnets are key functional materials for strategic emerging fields such as energy engineering and intelligent manufacturing. Misch metal (MM)‑based alloys help boost rare‑earth resource utilization, yet traditional trial‑and‑error experiments suffer from long cycles and cumbersome variable control. In this work, we established 18 model-optimization combinations by coupling six machine learning models with three hyperparameter optimization algorithms. A systematic test was conducted on datasets with different target properties, including Curie temperature Tc, residual magnetization Br, saturation magnetization Ms, coercivity Hc and maximum energy product (BH)max. For Tc prediction, 14 combinations achieved the coefficient of determination R2 > 0.9. For the small-sample datasets Br and Ms, 8 combinations maintained R2 > 0.79 with satisfying generalization capacity. The optimal combination, Extreme Gradient Boosting coupled with Grey Wolf Optimizer, was further applied to predict Hc and (BH)max of MMxFeyB100-x-y. The high-performance regions for Hc and (BH)max are x = 0.14 – 0.16, y = 0.79 – 0.80 and x = 0.125 – 0.13, y = 0.79 – 0.795. Its predictions closely matched experimental data, with average errors of merely 0.317 kOe and 0.016 MGOe. These findings provide guidance on the machine learning driven material design and performance optimization for permanent magnets.
High-resolution transmission electron microscopy (HRTEM) can reveal nanoscale structural heterogeneity, yet converting weak and ambiguous image contrast into reproducible multiclass structural maps remains difficult because reliable pixel-level annotations are scarce. Existing HRTEM segmentation workflows therefore often remain limited to binary foreground-background separation or other simplified tasks, rather than simultaneous parsing of crystalline regions, grain boundaries, amorphous or structurally unclear regions, and background. Here we report a physics-informed synthetic-to-experimental framework that enables few-shot multiclass HRTEM segmentation. The key advance is a Domain Construction-Relaxation-Imaging (DCRI) framework that generates synthetic HRTEM images with strictly co-registered structural labels, providing scalable supervision for synthetic pretraining. In this two-stage workflow, DCRI pretraining provides transferable structural priors, while few-shot experimental fine-tuning adapts these priors to real HRTEM images. Using Au nanoparticles as the primary model system, we show that this strategy produces coherent four-class segmentation under low-label conditions, whereas experimental-only training remains unstable. The benefit is retained across representative U-Net-family backbones and after few-shot adaptation to Au STEM images, shows partial transfer to annotated PbS HRTEM images, and further supports SAM-based foundation-model adaptation. These results establish a physics-informed synthetic-to-experimental framework as a practical route to alleviate the annotation bottleneck in few-shot structural segmentation of HRTEM images.
In this work, we present an efficient and fully ab initio approach for the calculation of the screened Coulomb potential, specifically designed for two-dimensional (2D) and three-dimensional (3D) metals and suitable to evaluate GW corrections. While quasiparticle renormalization effects are typically negligible in bulk metals, 2D metals may exhibit semiconductor-like features, such as sizeable renormalization of quasiparticle energies induced by Coulomb interaction, which follow from the suppression of screening due to the reduced dimensionality. Although this effect can in principle be described by the GW method, apart from a few pioneering exceptions, such calculations have been hampered by sampling requirements exceeding feasibility. We demonstrate that the combination of interpolation and extrapolation schemes together with the accounting of dynamical effects is pivotal to the correct handling of the intraband polarizability in the long-wavelength limit. As an application, we report the gap renormalization of selected doped 2D semiconductors in excellent agreement with ARPES measurements.
The fundamental nature of dynamic magnetoelectric coupling in multiferroics remains poorly understood. Despite these compelling static observations, its stability is governed by the convergence of the propagation direction (k) and the topology of the polarization structure, originating from the intertwining of π-disclinations. Strikingly, the knot exhibits synchronous propagation with ferroelectric domain walls under external electric fields, effectively preserving its topological integrity. Leveraging the topological nature of polarization order together with strong magnetoelectric coupling, a universal mechanism is manifested for the realization of nontrivial magnetic topological structures and their cooperative dynamics that should be operative across diverse topological materials. Our findings reveal a robust topology-mediated magnetoelectric dynamical mechanism, shedding light on the non-equilibrium physics of multiferroics and opening avenues for topological spintronics.
Artificial intelligence is transforming materials discovery from trial-and-error experimentation towards data driven design. Here we present a thermoelectric inverse-design (TEID) framework that integrates the MatHub-3d database, a generative model, and hierarchical screening across low-, medium-, and high-accuracy calculations. Trained on MatHub-3d, a crystal diffusion variational autoencoder generated 6400 candidate structures as the starting point of the workflow. Deduplication and stoichiometric/elemental constraints reduced this set to 1461 compositions. At the low-accuracy stage, pretrained machine-learning models rapidly narrowed the search space: M3GNet identified 1053 dynamically stable structures, and MEGNet returned 1005 thermodynamically viable compounds. At the density functional theory level, automated workflows comprising structural relaxation, convex hull analysis, band gap filtering, and deformation potential approximation electrical transport screening yielded 64 promising high-performance candidates. Phonon and third-order interatomic-force-constant calculations ultimately identified four dynamically stable and previously unreported compounds, DyTlTeSe, Li2In4Cu2Te8, Na3CaP3, and Si2Cu3AgS6 with lattice thermal conductivities of 0.67–1.72 W m−1 K−1. When combined with electrical transport calculations that include both deformation potential and polar optical phonon scattering, Na3CaP3 and Si2Cu3AgS6 achieve exceptional n-type performance, with ZT values of 1.59 and 2.24 at 700 K. These results demonstrate TEID as a generalizable framework for data-driven discovery and validation of novel functional materials.
The magnetocaloric effect (MCE) offers compact low-temperature refrigeration for quantum and energy technologies, but practical use requires affordable, synthesizable materials with high performance. Here, first-principles calculations and atomistic spin simulations identify halogen substitution as an effective route to tune two-dimensional transition-metal phthalocyanine frameworks (M-XPc; M = Cr, Mn; X = F, Cl, Br, I). Halogen atoms induce lattice expansion and charge redistribution, driving antiferromagnetic-to-ferromagnetic transitions and modifying magnetic anisotropy. This chemical control produces inverse, rotating, and conventional MCE within one rare-earth-free material family. Cr-BrPc shows a large inverse MCE, with a maximum magnetic entropy change of –9.04 µJ m−2 K−1 (–6.21 J kg−1 K−1) under a 5 T field change, and a rotating entropy change of 8.73 µJ m−2 K−1 (5.91 J kg−1 K−1). Mn-IPc exhibits enhanced conventional MCE near the hydrogen-liquefaction temperature range, reaching 5.47 µJ m−2 K−1 (2.88 J kg−1 K−1) and an adiabatic temperature change of 4.65 K. These results establish halogenated MPc frameworks as tunable rare-earth-free candidates for cryogenic magnetic refrigeration.
Sliding ferroelectricity in van der Waals layers presents a novel mechanism for polarization switching. It features an ultralow energy barrier and high endurance, both governed by domain wall (DW) motion. Compared to bilayer sliding ferroelectric systems, trilayer and multilayer architectures introduce multiple sliding interfaces, causing polarization switching to involve intermediate domain states and coupled DW dynamics rather than a simple single-interface process. Nonetheless, the realization of these multistate switching functionalities is strongly dependent on DWs dynamics, which become substantially more intricate in these layered architectures and are still poorly understood. Here, using trilayer γ-InSe as a minimal prototype of multilayer sliding ferroelectrics, we systematically investigate the static and dynamic properties of DWs by combining density functional theory calculations and machine-learning-assisted molecular dynamics simulations. The simulation results reveal that adjacent layers form DWs at distinct locations, separated by nonpolar domains situated between the upward- and downward-polarized ferroelectric domains, which markedly differs from the bilayer structure. Under external electric fields, polarization switching proceeded through distinct kinetic stages, with DW motion exhibiting a parabolic displacement-time relationship that transitioned from acceleration to DW velocity at higher fields. Notably, the collective motion of the DWs indicates the presence of significant inter-DW interactions. This work provides an atomistic picture of intermediate nonpolar domains and staged DW-mediated switching dynamics in trilayer sliding ferroelectric γ-InSe, providing microscopic insight into DW-mediated switching in multilayer sliding ferroelectrics.
FePt-X nanocomposite thin film (X is a non-magnetic segregant) is a critical enabling material for ultrahigh-capacity magnetic recording, but a time- and cost-efficient strategy to identify process parameters that yield desirable microstructure features has remained scarce. Here, we present a computational framework, which employs high-throughput phase-field simulations for the forward prediction of process-microstructure relationship and integrates automated microstructure quantification and active learning for the inverse identification of process parameters. As a proof of concept, the framework is used to inversely identify the gradient energy coefficients that lead to microstructure features matching experimental observation in FePt-Al2O3 films. The identified parameters result in a closer agreement with experiments than expert best guesses. Notably, the framework yields a tenfold reduction in computational cost compared with brute-force high-throughput phase-field simulations. Our results demonstrate the potential of integrating phase-field simulations with active learning to significantly accelerate inverse process design for microstructure engineering, compared with using phase-field simulations alone.
High-entropy MBenes (HE-MBenes) are two-dimensional transition metal borides whose metal sublattice is occupied by five or more principal elements, forming a largely unexplored class of electrocatalysts for CO2 electroreduction. These materials combine the lone-pair-driven CO2 activation intrinsic to MBenes with the configurational stabilization and chemical diversity of high-entropy systems, circumventing the linear scaling relationships (LSRs) that constrain monometallic surfaces. We report a multi-fidelity computational screening of all 56 equiatomic quinary HE-MBene compositions from a {Ti, V, Cr, Mo, Nb, Ta, Zr, Hf} elemental pool. A sequential funnel integrating DFT relaxation, formation-energy filtering, PDOS-guided active-site identification, and computational hydrogen electrode (CHE) free-energy profiling, accelerated by a MACE machine-learning interatomic potential across 1375 surface environments, narrows the space from 56 to 18 viable catalysts. Three compositions, CrNbZrMoTiB5, MoZrHfNbCrB5, and MoZrHfTaCrB5, exhibit spontaneously downhill CO2-to-CO free-energy profiles at zero applied potential (UL = 0.00 V vs. RHE), surpassing all previously reported CHE-computed limiting potentials for MBene and high-entropy-alloy catalysts. Bader charge analysis reveals that Hf/Zr-mediated electron donation to Cr-centred active sites stabilizes the rate-determining COOH* intermediate, decoupling LSRs. These findings establish HE-MBenes as a generalizable catalyst platform and provide a transferable multi-fidelity framework for multicomponent 2D materials discovery.
Abstract First-principles defect calculations in hematite ( α -Fe 2 O 3 ) suffer from a >2 eV spread in reported formation energies, arising from (i) metastable or delocalized charge states and (ii) inconsistent error cancellation with chemical-potential boundaries across studies. We introduce a charge-localization protocol based on the Fe-3 d occupation matrix that enforces physically meaningful small-polaron configurations and eliminates the spurious loss of intermediate charge states; the resulting thermodynamic charge-transition levels are essentially independent of computational parameters. To establish a common energy scale, we use the crossing point of the fully ionized Frenkel pair ( $${V}_{{\rm{Fe}}}^{3-}$$ V Fe 3 − and $${{\rm{Fe}}}_{i}^{3+}$$ Fe i 3 + ) as an intrinsic anchor and align the chemical potentials to parameter-independent experimental formation enthalpies. This reduces the inter-study deviation by ~80%, and reveals that the remaining absolute shifts are dominated by VBM displacement amplified by the large nominal defect charges. The framework provides a reproducible benchmark for hematite defect energetics and is directly transferable to other correlated oxides.
Reliable uncertainty quantification (UQ) for graph neural networks (GNNs) under out-of-distribution (OOD) shifts remains insufficiently characterized in materials discovery. Existing benchmarks based on random splits can overestimate model reliability by underrepresenting structural extrapolation challenges. Here we introduce MatUQ, a benchmark built on structure-aware Smooth Overlap of Atomic Positions Leave-One-Cluster-Out (SOAP-LOCO) splitting, together with a training protocol that combines Deep Evidential Regression (DER) with dropout regularization, for evaluating GNN reliability under structural distribution shifts. Through systematic experiments spanning six materials datasets, twelve GNN architectures, and eight UQ strategies, we find that predictive accuracy and uncertainty quality are distinct capabilities that tend to decouple under OOD evaluation, and that uncertainty-metric leadership is largely non-transferable across datasets and target properties. We further find that as training data become scarce, the optimal strategy shifts from evidential-containing hybrids toward pure ensemble variance across all evaluated architectures, with the architecture holding the distributional optimum shifting correspondingly. Standalone evidential regression rarely attains per-model optima and benefits from hybrid pairing only under specific combinations of data density, inductive bias, and target metric. Monte Carlo dropout is less effective as a standalone uncertainty estimator under the tested configurations but can contribute within hybrid schemes. Overall, MatUQ provides a more rigorous benchmark for assessing uncertainty-aware GNNs in OOD materials discovery.
Membrane-based separation is a critical technology for energy and environmental processes, positioning it at the forefront of materials science research. To facilitate the data-driven advances in this field, we developed an integrated natural language processing pipeline including a domain-tailored transformer model for literature mining. One major challenge is data imbalance which biases the model toward majority classes and leads to overfitting on dominant patterns. We introduced two complementary strategies to direct the model’s attention toward minority classes: (1) a data augmentation method that prevents data leakage, and (2) model fine-tuning using a customized loss function that integrates class-weighted cross-entropy and focal loss. The resulting model MembraneBERT achieves an average Precision of 0.906, Recall of 0.796, and F1-score of 0.844, respectively. Applied to 3,580 additional articles, the model extracted 2,283 structured entries on membranes and their gas separation performance, offering a practical framework for AI-driven extraction of membrane performance data to support subsequent data-driven research.
Ultra-wide bandgap (UWBG) semiconductors host the large critical electric fields necessary to support a leap in the power density of next-generation power and radio-frequency electronics. However, effective thermal management of the heat flux incited by high-voltage operation remains a limiting challenge. The current understanding of thermal transport in UWBG semiconductors neglects the departure from equilibrium driven by electric field ($$\mathop{E}\limits^{ \rightharpoonup })$$, which becomes more severe at higher power density. In this work, we develop a new computational workflow to obtain non-equilibrium ab initio thermal transport properties under $$\mathop{E}\limits^{ \rightharpoonup }$$ within the density-functional theory - non-equilibrium Green’s function (DFT-NEGF) framework. The predicted cross-plane thermal conductivity (k⊥) of wurtzite Aluminum Nitride (AlN) is modified by a cross-plane electric field ($${\mathop{E}\limits^{ \rightharpoonup }}_{z}$$), where the change in k⊥ depends on the direction of $${\mathop{E}\limits^{ \rightharpoonup }}_{z}$$ with respect to the polarity of the AlN thin film. The findings provide critical insights to understand thermal transport in devices under operating conditions and design effective thermal management for high-power electronics.
Although antiferromagnetic (AFM) materials surpass conventional ferromagnetic materials in multiferroic tunnel junctions (MFTJs) due to their ultrafast magnetodynamic response, their application is limited by the difficulty in reading the tunneling magnetoresistance using conventional methods. To address this challenge, we propose breaking the PT symmetry of bilayer FeOCl by vertically flipping. This symmetry-breaking induces spin-splitting in AFM bilayer FeOCl. We then design fully compensated ferrimagnet bilayer FeOCl and antiferroelectric bilayer In2Se3-based tunnel junctions. Using first-principles calculations combined with the non-equilibrium Green’s function method, we demonstrate that various non-volatile resistance states can be realized by adjusting the magnetic state of electrodes and the polarization direction of the ferroelectric barrier. The devices show a high tunneling magnetoresistance (TMR) of 1.8 × 103% and electroresistance (TER) of 1.1 × 105%. This study presents a method to induce spin-splitting in two-dimensional AFM materials and highlights their applications in high-speed and high-density AFM-based spintronic devices.
Coupling multiferroic orders with emerging altermagnetism offers a promising route toward multifunctional quantum devices. Here, we combine symmetry analysis, systematic density functional theory screening, and nonequilibrium Green’s function calculations to investigate 48 NiAs-type (110) monolayers. We identify a class of ferroelastic-altermagnetic biferroics exhibiting intrinsic coupling between lattice strain and spin order. Specifically, metallic FeN and semiconducting MnTe monolayers emerge as representative systems, showing reversible ferroelastic switching with low barriers of 0.08 and 0.22 eV per formula unit and accessible signal intensity. FeN exhibits above-room-temperature (334 K) altermagnetism while MnTe possesses Néel temperature of 122 K. Both monolayers display ferroelastic-reversible momentum-dependent spin splitting (245 and 307 meV). This coupling enables tunable anomalous Hall conductivity and strain-controlled electronic transport. Prototype device-level simulations reveal magnetoresistance (454%), elastoresistance (155%), and high spin-filtering efficiency (-87%). These findings establish NiAs-type monolayers as a versatile platform for intrinsic biferroic coupling, providing a foundation for high-density, energy-efficient spintronic architectures.
Aqueous lithium-ion batteries provide a promising route toward safe and sustainable energy storage, yet their energy density is constrained by the narrow electrochemical stability window of aqueous electrolytes. Although high salt concentrations can kinetically extend the cathodic limit of the electrolyte via anion-derived interphases, this effect depends strongly on the electrode material and is absent on platinum. To elucidate the interfacial reaction mechanisms behind the absence of cathodic-limit extension on platinum, we performed molecular dynamics simulations based on a machine learning force field, combining near first-principles accuracy with large-scale configurational sampling. The simulations reveal a dominant pathway for anion decomposition on a hydrogen-covered platinum surface under cathodic conditions, mediated by adsorbed hydrogen (H*). However, this pathway directly competes with hydrogen evolution via the Tafel step (2H* → H2) at comparable activation barriers, suppressing the formation of stable anion-derived interphases and the extension of the cathodic limit. These findings reveal interfacial H* coverage as a key factor governing electrode-dependent cathodic stabilization in concentrated aqueous electrolytes.
Compositional tuning between materials with contrasting magnetic and electronic properties can induce behaviors absent in their end members. In this paper, we demonstrate that in mixed MnTe-MnSb-MnBi alloy systems, the competition between MnTe, an antiferromagnetic (AFM) semiconductor with altermagnetic band symmetry, and the ferromagnetic (FM) metals MnSb and MnBi drives several composition-dependent transitions. Using first-principles calculations of disordered supercells we reveal that an AFM-FM boundary near MnTe0.75Sb0.25 exhibits a pronounced magnetoelastic response, disorder-induced phonon broadening, and a semiconductor-to-metal crossover. These effects are associated with enhanced fluctuations in local bonding and interatomic force constants near the AFM-FM crossover. This work elucidates how magnetic competition and local chemical disorder shape the structural, vibrational, and electronic behaviors of MnTexSbyBi1-x-y alloys, thereby driving large magnetoelastic responses.
Efficient multi-objective collaborative design is a core scientific challenge and a primary obstacle to the development of novel glass materials. Here, we propose IGlasGAN (Inorganic Glass Generative Adversarial Network), a generative artificial intelligence framework for the automatic inverse design of inorganic glass materials subject to box constraints on multiple coupled properties, alongside a full generation-evaluation-validation pipeline. Built upon an improved WGAN-GP-CP architecture, the framework incorporates a property navigation mechanism and an extended normalization strategy to enable end-to-end, on-demand design, from property requirements to glass compositions, across a broad design space. By training on a dataset comprising 2445 experimental samples, in which 156 samples fully satisfy the four property requirements, including coefficient of thermal expansion (CTE), Young’s modulus (E), strain point temperature (Tst), and density (ρ), a multi-objective inverse design for the OLED substrate glass utilizing IGlasGAN was carried out, and novel compositions with relatively high Y₂O₃ content was identified. Two additional design-from-scratch scenarios with progressively more distant target regions from the training data distribution were constructed to further validate the framework’s generation capability. Across the three scenarios, 12 generated samples were experimentally characterized. The measured properties agreed well with the generated values, with deviations within ±3.50%, except for CTE, which showed a maximum deviation of 13.03%. These results confirm the applicability of the IGlasGAN framework across diverse target settings, providing an efficient intelligent pathway for multi-objective inverse design of inorganic oxide glasses.
