Traditional metasurface inverse design methods suffer from intensive manual modeling, laborious data collection, and extensive numerical calculations, leading to a compromised design efficiency and accuracy. To solve these drawbacks, this paper proposes a metasurface inverse design method using a conditional generative adversarial network (CGAN). A dataset construction framework integrating modeling, simulation, and response processing is established to ease data samples collection and reduce manual workload. Different from conventional parameterized modeling, the metasurface unit patterns are constructed by using a pixel‐based graphical modeling method based on encoding matrices. Furthermore, Gramian Angle Fields (GAFs) are incorporated to explore the latent physical correlations in electromagnetic responses, effectively improving the quality of generated metasurface structure images. By embedding an optimized U‐Net architecture into the CGAN model to suppress gradient vanishing and improve training stability and resolution, high‐precision mapping from target responses to metasurface structures is realized. An electromagnetically induced transparency (EIT) metasurface validates the proposed method. The results demonstrate that the proposed method achieves lower MSE, fewer training epochs, and less training cost. The simulated transmission spectra of the designed metasurface agree well with the target spectra with an average spectral accuracy above 90%, exhibiting promising application potentials for metasurfaces inverse design.
Advanced tactile skin technologies capable of simultaneously measuring contact location and multiaxis forces are essential for next‐generation human–robot interaction and dexterous manipulation. However, existing sensors remain limited in simultaneously resolving multidimensional signals under combined loading. Here, we present a self‐assembled magnetic cilia‐based soft skin sensor with a hybrid decoding pipeline for real‐time five‐degree‐of‐freedom (5‐DoF) tactile estimation. A magnetic field‐induced self‐assembly process yields high‐aspect‐ratio cilia arrays without lithographic patterning, achieving a signal‐to‐noise ratio (SNR) improvement of approximately 24 dB over planar magnetic films. A hybrid decoding pipeline that combines a convolutional neural network (CNN) with a calibrated physics‐guided shear model incorporating Hertzian contact‐area scaling addresses the scale imbalance between normal and shear force signals, achieving a localization error of 0.47 mm and force estimation errors below 8%. The system is demonstrated through teleoperation of a six‐degree‐of‐freedom (6‐DoF) collaborative robot arm, illustrating its feasibility as a tactile command interface.
Precise tuning of physical properties, such as lower critical solution temperature (LCST) behavior in polymer formulations, is essential for stimuli-responsive materials design, but is typically limited by low-throughput cloud point (T cp) determination. To address this, we developed a high-throughput (HTP) experimentation platform that integrates automated liquid handling, parallel turbidimetry measurements, and machine learning (ML) under a unified Python framework. Using an OT-2 liquid handling robot interfaced with a SPECTROstar Nano plate reader, we enabled automated design of experiments-based sampling of high-dimensional formulations, which are prepared robotically and analyzed via HTP turbidimetry to determine polymer cloud points (T cp) for P(NIPAM-co-DMAEMA) copolymers. Overcoming single-variable resolution in traditional methods, our approach enables the simultaneous T cp mapping across polymer, NaCl, and NaOH concentrations, achieving an average measurement throughput of 8 min per sample. This workflow generates high-fidelity datasets comprising hundreds of formulation-T cp relationships within days. Supervised ML models trained on these data achieved high predictive accuracy (R 2 up to 0.96; mean squared error as low as 0.023 degrees C2), with Shapley Additive Explanations analyses elucidating feature contributions. T cp distribution analysis identified polymer concentration as a dominant driver of T cp variance, particularly in hydrophilic regimes. Using ML, inverse design of formulations was demonstrated, predicting compositions that achieve target transition temperatures with less than 1% error. By combining automated formulation, HTP cloud point measurement, and ML in one workflow, our platform addresses the asymmetry in throughput between polymer synthesis and physical property characterization that currently limits data-driven LCST design.
In tightly cooperative manipulation tasks, robotic manipulators must follow collision-free and coordinated trajectories. Existing multiagent learning frameworks often rely on centralized planners that provide strong coordination but fail to scale with larger teams. Alternatively, decentralized approaches offer better scalability but typically lack communication, which limits their ability to achieve highly cooperative behaviors. To address this gap, this paper proposes a graph neural network (GNN)-based framework for scalable multiagent reinforcement learning (RL). In our formulation, each manipulator is represented as a node in a GNN, and message-passing edges provide a communication mechanism that enables agents to share information effectively. This design allows the team to achieve flexible decentralized planning while maintaining strong cooperation. Our results also demonstrate the scalability of the approach, showing that a single control policy can be trained once and successfully applied to tightly cooperative manipulation tasks across teams of varying sizes, without retraining.
Raman spectroscopy is a widely used tool for nanoscale materials characterization, yet weak spectral features are often obscured by strong and spatially variable background signals. This challenge is particularly severe in interfacial and low-dimensional systems, where dominant substrate responses make conventional reference-based subtraction unreliable. Here, we introduce a transformer-based deep learning framework for reference-free spectral unmixing that reconstructs substrate contributions directly from partially observed spectra. By exploiting self-attention mechanisms to capture nonlocal spectral correlations, the model learns complex background signatures without requiring dedicated reference measurements. Subtraction of the reconstructed background enables the recovery of weak, previously inaccessible spectral features. We demonstrate the approach on buffer layer graphene grown on silicon carbide, a prototypical background-dominated system, where the model reveals vibrational signatures of the buffer layer otherwise hidden by the substrate response. The extracted features are validated against ab initio calculations, confirming their physical origin. Beyond this specific case, the framework provides a generalizable strategy for robust, automated spectral unmixing, compatible with real-time acquisition and closed-loop, artificial intelligence-assisted experimental workflows.
