Large Language Model chatbots have gained significant popularity, offering knowledge to support specialists in diverse fields. However, adapting models to specific use cases and specialized domains presents considerable challenges. Hence, we introduce the LLM Playground, a comprehensive approach to optimizing LLMs for specialist applications with respect to their accuracy in answering domain‐specific questions, addressing the limitations of unmodified models. The utilized optimization techniques begin with Prompt Engineering, advance to the integration of external knowledge, and culminate in complex reasoning strategies or self‐feedback loops. This paper introduces various architectures for scientific assistants, comprising individual enhancement techniques, both in isolation and in combination with others, designed to facilitate comparisons. To demonstrate the efficacy of the LLM Playground, a chemical chatbot is set up as a case study, and the optimization techniques are compared using ChemBench, an independent question–answer benchmark for the chemical domain, to measure its performance. By providing tested, ready‐to‐deploy architectures and clear use‐case guidance, this work helps researchers and practitioners leverage LLMs in domain‐specific applications. The insights and methodologies presented in this paper contribute to the growing body of knowledge on tailoring LLMs to meet the unique demands of specialized fields.
The biomedical domain's accelerating progress in understanding, early detection, and treatment of diseases has created an exponentially growing and overwhelming body of literature. Researchers rely on this literature to find relevant information, but navigating this vast landscape has become increasingly challenging, especially for interdisciplinary AI‐biomedicine researchers who need to stay current across both highly fast‐paced fields. Despite the emergence of large language models (LLM) systems, retrieving precise, domain‐specific literature remains a significant challenge. This paper addresses these challenges by integrating knowledge graphs with scientific literature embedded in large language models to expedite biomedical discovery. We employ a novel strategy to discover the most relevant pathways between biomedical entities in knowledge graphs. These pathways are then leveraged by a multiagent LLM system to formulate facts from literature, design AI predictors for understanding discovered pathways, and propose wet‐lab experiments to validate AI predictions. This approach creates a comprehensive end‐to‐end methodology for biomedical discovery. Experiments with various biomedical entity pairs demonstrate the framework's ability to identify highly relevant pathways and design plausible, complex AI predictors with wet lab validation experiments across diverse therapeutic areas. We developed Intelliscope, a first‐of‐its‐kind platform to accelerate scientific discoveries, potentially leading to breakthroughs in disease understanding, drug repurposing, and therapeutic development.
Current experimental workflows in solution‐processed thin‐film photovoltaics research are restricted to known solution combinations where solvent properties are predictable. To experimentally investigate the vast parameter space of new compounds, we must explore unknown material combinations using automated systems. However, this increases the risk of coating failures that previously required human judgment to detect. Integrating real‐time process monitoring into existing deposition equipment is essential to close that gap. We present a vision‐based add‐on to commercial spin coaters that provides the process feedback missing from standard equipment—wetting assessment and substrate orientation for robotic retrieval—enabling fully automated coating workflows. Our hybrid software pipeline combines semantic segmentation using a width‐scaled U‐Net with classical geometric analysis, enabling simultaneous contact angle measurement and substrate pose estimation from a single RGB camera. We validated the module through an 11 h fully automated experiment, tracking hydrophobic recovery on plasma‐treated substrates. By replacing human process supervision with machine vision, this system is a prerequisite for materials discovery in previously unexplored chemical regimes and transforms standard commercial spin coaters into components of Self‐Driving Labs.
We demonstrate a single SiO x ‐based threshold switching (TS) device for dual‐mode operation, such that full‐oscillation and probabilistic‐oscillation (p‐oscillation) can be controllably selected using the input voltage ( V in ) range of 2–6 V. This functionality relies on the TS mechanism, involving the formation and spontaneous dissolution of a weak conductive filament. By optimizing the Ar:O 2 gas flow ratio during SiO x deposition, we modulate the oxygen vacancy concentration, determining the degree of variability in filament formation. This device exhibits uniform TS characteristics under low V in ( V in ≤ 3.5 V). When an AC pulse is applied to the device, voltage spikes are periodically generated owing to charging and discharging cycles, exhibiting stable full‐voltage oscillation. These devices serve as an oscillation neuron in oscillatory neural networks to accurately recognize noisy patterns. Under high V in ( V in ≥ 3.5 V), enhanced instability of these devices induces intermittent oscillation failures, yielding p‐oscillation. The probability of these spikes follows a V in ‐controllable sigmoid distribution, effectively leveraged as a probabilistic bit to solve vehicle routing problems. We implement this p‐oscillation as a robust entropy source for a true random number generator by integrating a five‐input XOR postprocessing circuit. This architecture provides critical immunity to probability deviations and ensures that the generated random numbers pass all 15 National Institute of Standards and Technology SP 800‐22 tests, as validated via physical model‐based simulations.
Organic molecular crystals are critical for many pharmaceutical and functional materials. However, their discovery is hindered by the combinatorial complexity of molecular design, polymorphic packing, and crystallization conditions. This review describes how materials informatics and autonomous experimentation can accelerate discovery through an integrated molecule–crystal–function–optimization workflow. Molecular‐level structure–property modeling and machine learning (ML) for the low‐cost screening of large chemical spaces are outlined. Crystal‐level approaches for predicting crystal stability and functional properties, including machine learning interatomic potentials (MLIPs) and finite‐temperature molecular dynamics, are discussed. Crystal structure prediction (CSP) is reviewed as the key link between molecules and realizable crystal packing, emphasizing ML‐accelerated structure generation, stability ranking beyond 0 K lattice energies, and strategies for managing polymorph overprediction. Bayesian optimization (BO) and closed‐loop robotic platforms that couple computation with crystallization and characterization to enable autonomous laboratories are highlighted. Open challenges include data quality and negative‐result collection, MLIP transferability, BO scalability to high‐dimensional and multi‐objective settings, and standardized benchmarks needed to transform data‐driven organic crystal discovery into a practical pipeline. The ultimate vision is a prompt‐to‐materials autonomous discovery engine, where humans define objectives and then ML, crystal prediction, and robotic workflows iteratively optimize discoveries.
In high‐throughput applications, crystallographic analysis via X‐ray diffraction (XRD) is often limited by long exposure times and the need for manual data interpretation. This study presents a novel machine‐learning‐based approach for volume fraction estimation from XRD patterns addressing both challenges. The method efficiently processes noisy XRD patterns acquired with polychromatic emission spectra, and enables volume fraction estimation from thousands of patterns per second. Compared to conventional techniques, it requires much lower XRD pattern quality, allowing for shorter exposure times. This makes our method particularly suited for high‐throughput scenarios such as self‐driving labs. We utilize simulated XRD patterns to train two neural networks to process XRD data in a specified material system. A convolutional neural network (CNN) estimates volume fractions, while a u‐net‐style network restores pattern interpretability by resolving peak duplication caused by polychromatic emission spectra. We use synthetic datasets to showcase the method's noise tolerance and ability to analyze XRD patterns with multiple emission lines. Furthermore, we verify our method's applicability to real XRD patterns using a small experimental dataset.
The growing complexity of thermal power generation systems demands advanced forecasting solutions capable of integrating data analysis, model selection, and interpretability. This study proposes a modular large language model (LLM) agent framework for time series forecasting, designed to operate locally and interactively through natural language instructions. The framework incorporates a domain‐specific time series agent that was developed to automate data preprocessing, anomaly detection, and forecasting tasks using neural and statistical models. Experiments demonstrated the agent's capacity to autonomously conduct end‐to‐end analyses, achieving accurate forecasts with minimal user intervention. The PatchTST model, automatically selected by the agent, yielded the lowest mean squared error among evaluated methods. Results highlight the potential of LLM‐based agents to enhance transparency, usability, and reproducibility in energy forecasting pipelines.
Molecular‐level visualization of ion transport and separation dynamics in complex environments is crucial for advancing energy systems, water purification, and critical materials recovery. Achieving this requires imaging platforms that combine structural sensitivity, chemical specificity, and real‐time operation. Digital off‐axis holography (DOAH) provides high‐throughput, label‐free quantitative phase imaging but inherently lacks chemical selectivity. Integrating DOAH with complementary spectroscopic channels such as fluorescence or hyperspectral imaging introduces the needed molecular specificity, while also creating challenges in multimodal data fusion, synchronization, and computational throughput. Artificial intelligence (AI) offers a powerful route to address these limitations by uniting physics‐based reconstruction with data‐driven interpretation. In this perspective, we outline a framework for intelligent multimodal holography and demonstrate its potential using a preliminary AI‐driven test case. Raw DOAH holograms of lanthanide solutions subjected to magnetic field gradients were analyzed using multi‐agent AI workflows that autonomously selected reconstruction tools, extracted NMF components, and generated scientific claims consistent with the expected paramagnetic and diamagnetic behavior. This demonstration shows how AI‐enabled reasoning can deliver real‐time chemical–structural interpretation directly from raw holograms. Together, these advances define a path toward adaptive, intelligent holography platforms capable of supporting in situ chemical separations, dynamic ion transport analysis, and next‐generation interfacial science.
Traditional network pharmacology (NP) is fundamentally constrained by its reliance on static, two‐dimensional topological mapping and pervasive data sparsity, limiting its capacity to capture the dynamic complexity of biological systems. To transcend these descriptive boundaries, we propose and operationalize AI‐NP as a rigorous, multimodal computational framework that reconceptualizes biological networks as dynamic, learnable manifolds rather than fixed graphs. The paradigm is organized around four core methodological pillars: knowledge graph engineering, geometric graph learning, multimodal state anchoring, and generative reasoning paradigms. By integrating these pillars, AI‐NP enables the mechanistic decoding of complex polypharmacology inherent in multicomponent herbal formulas. Leveraging state‐of‐the‐art architectures, including graph attention‐aware fusion networks for zero‐shot link prediction and SO(3)‐equivariant message‐passing models for 3D stereochemical resolution, AI‐NP establishes a computationally rigorous pipeline that transitions the field from empirical observation to mechanistic prediction. This paradigm shift offers a scalable and scientifically validated pathway for discovering effective, equitable therapeutics from traditional medicine systems.
In order to help in the better application of seawater, the hydration behavior of tricalcium silicate (C 3 S) was studied. C 3 S was mixed with seawater‐relevant solutions (NaCl, MgCl 2 , Na 2 SO 4 ) with varied concentrations. The simplified C 3 S model system was adopted to systematically isolate the roles of individual seawater ions and their interactions. The results showed that seawater‐relevant solutions increased the hydration rate of C 3 S at an early age. The heat flow peak value increased with increasing salt concentration. Among the investigated salts, MgCl 2 exhibits the most obvious effect, indicating its dominant role in the acceleration behavior of seawater. The effect of seawater was more significant than any single salt solution with its typical concentration in seawater. The nucleation rate of hydration products was increased with the addition of salts, which was the main reason for the acceleration effect. A neural network was introduced to help understand the effect of seawater relevant solutions on the simplified model system. The neural network was trained based on the experimental data. The predicted calorimetry results agreed with the experimental data. The average coefficient of determination of the model exceeds 95%, which highlights its potential as a practical tool for predicting hydration behavior from seawater composition.
Cell segmentation is a cornerstone of biological image analysis, enabling downstream applications such as tracking, counting, and phenotypic profiling. Deep learning has markedly improved segmentation accuracy, leading to broad adoption in research and clinical workflows. However, most current approaches operate on 2D images, neglecting spatial information in the third dimension. This omission is critical, as cellular morphology, migration, and division are inherently three‐dimensional processes. Modern microscopy now offers high‐throughput volumetric imaging, presenting new opportunities for methods that fully exploit 3D data. This review surveys segmentation strategies that incorporate depth information, extending beyond conventional 2D analysis. It distinguishes between pseudo‐3D (2.5D) methods, which balance computational efficiency with limited volumetric context, and fully 3D methods that process volumetric input and output directly. In total, 31 beyond‐2D segmentation approaches are examined and compared, alongside 32 curated volumetric datasets annotated with key metadata such as resolution, size, and ground‐truth availability. Furthermore, a unified reference dataset, compiled from high‐quality open‐source resources, is presented to promote standardized data access, metadata harmonization, and future community‐driven benchmarking efforts in 3D cell segmentation.
The exponential proliferation of cancer multi‐omics data offers unprecedented opportunities for precision oncology but presents severe challenges due to high dimensionality, sparsity, and heterogeneity. This article systematically reviews how artificial intelligence (AI) reshapes this landscape. First, we explore generative models (Variational Autoencoders, Generative Adversarial Networks) that reconstruct high‐fidelity data foundations through intelligent imputation and non‐linear batch correction. Second, we dissect the evolution of integration architectures, highlighting Graph Neural Networks for topological feature extraction and Transformers for constructing biological foundation models. Addressing clinical trust, we evaluate Explainable AI (XAI) strategies for transparent biomarker discovery. At the clinical level, we demonstrate AI's superior efficacy in molecular subtyping, survival stratification, and immunotherapy prediction. Finally, we identify meta‐learning, spatial multi‐omics, and federated learning as pivotal directions for realizing the clinical translation of next‐generation precision oncology.
This study evaluates the efficacy of three foundation potentials (FPs)—SevenNet, DPA, and Orb—in predicting selected properties of several alloys and metals that are based on three elements: Co, Ni, and Ru. The analyzed systems comprise the face‐centered cubic (FCC) structured CoNiRu, Co 2 Ni 2 Ru, and pure Ni, as well as hexagonal close‐packed (HCP) structured CoNiRu and CoNiRu 2 . Properties include lattice parameters, elastic constants, and generalized stacking fault energies (GSFEs), assessed at 0 and/or 300 K. Results are compared against density functional theory (DFT) data whenever available. We found that all FPs typically forecast lattice parameters within 2.5% of DFT values at 0 K. At 300 K; both DPA and SevenNet largely successfully capture the expected thermal expansion trend. Orb was excluded at 300 K due to lattice instability. For elastic constants, DPA consistently captures thermal softening at elevated temperatures, while Orb and SevenNet exhibit potential‐dependent discrepancies. For GSFEs, DPA exhibits the closest alignment with DFT in both FCC and HCP structures, while Orb consistently shows the poorest performance. By elucidating the respective strengths and limitations of FPs, our findings offer insights into the robustness of these interatomic potentials for high‐throughput screening and predictive modeling of alloys.
Monitoring real‐time interactions in microrobotic and biological systems necessitates rapid multicolor fluorescence microscopy to resolve cellular, micro‐agent, and environmental dynamics. Sequential acquisition is widely used to capture multiple fluorescence channels while minimizing spectral crosstalk and photobleaching. However, this approach limits imaging speed in proportion to the number of channels, reducing its suitability for capturing micro‐agents and cellular interactions. This study introduces a real‐time multicolor reconstruction framework that exploits cross‐channel inputs (frames containing mixed spectral contributions) to enable simultaneous reconstruction of target fluorescence channels. The framework is evaluated on a custom‐built three‐channel fluorescence microscope, benchmarking two representative models: a standard supervised convolutional encoder–decoder with skip connections (U‐Net) and an adversarially trained conditional model (pix2pix). Experimental validation is conducted in a microfluidic environment containing active functionalized structures (Coumarin 6‐labeled electrospun magnetic fibers) and passive biological agents (CellTracker Red CMTPX‐labeled HeLa cell spheroids), where external magnetic fields actuate the micro‐agents to drive interactions with the spheroids. Prediction performance is evaluated in two‐ and three‐channel settings, yielding high‐fidelity reconstructions at 13.9 ms inference time. The proposed approach can increase the effective frame rate for three‐color channels by up to 83%, enabling high‐throughput multicolor imaging.
Exploration of polymer materials using machine learning (ML) has been actively demonstrated for wide applications, while the multiple‐purpose ML and property prediction using small data is challenging. Herein, we propose a low‐cost refractive index prediction model of polymers using a multiple regression model with domain knowledge and variance inflation factor, enabling higher prediction accuracy (MAE = 0.03) and extrapolability even with a small dataset ( n ~ 40). Through combining this prediction model with the multi‐objective Bayesian optimization (MOBO) and the generative model, several polymer candidates with both high refractive index ( n D > 1.75) and visible light transparency are predicted and experimentally demonstrated. Overall, this work provides a less time‐/cost‐consuming approach for the accurate property prediction and the simultaneous maximization of multiple properties in polymers, even if only a small‐scale database is available.
Multicomponent oxides are important materials for energy, catalytic, and electronic applications, but their huge composition space makes exploration by conventional solid‐state synthesis inefficient. Machine learning‐based selection is combined with a two‐step experimental workflow for pseudo‐ternary oxide discovery. Pseudo‐ternary systems with no registered compositions in an Inorganic Crystal Structure Database are ranked by the average prediction score of a model trained only on pseudo‐binary oxides, and the top 300 systems are screened by a slurry‐based solid‐state reaction method suited for robotic dispensing. Among the 300 tested compositions, 21 show diffraction peaks that cannot be explained only by simple or binary oxides. Three match ternary oxide phases in a powder diffraction database, 15 suggest cation‐disordered ternary phases, and two copper‐containing samples form oxides with copper valence states different from those assumed in the prediction. One remaining composition yields a new oxide, Ba 5 SnV 6 O 22 . Structural analysis shows that this phase has a Ba 2 BiV 3 O 11 ‐related framework, in which the Bi‐equivalent site is randomly occupied by Sn and Ba. These results show that prediction‐guided selection and a robotics‐compatible synthesis workflow can work together for oxide discovery.
