thermodynamic modeling

This work addresses the limitation of current large language models (LLMs) in multi-zone HVAC control, which typically lack explicit modeling of building physics and thermodynamic processes. To bridge this gap, the authors propose a knowledge graph that integrates thermodynamic principles with spatial semantics, constructed upon the Brick ontology and enriched with historical environment-controller interaction data to provide LLMs with structured contextual information. This approach represents the first integration of physics-informed spatial semantic graphs into an LLM-based control framework, explicitly capturing inter-zone thermal couplings and building dynamic responses. Evaluated in a five-zone building simulation, the method significantly improves the trade-off between energy efficiency and occupant comfort compared to both conventional and existing LLM-based strategies, achieving the lowest PMV violation rate while maintaining high energy performance.

This work addresses the challenge of training thermodynamic computing hardware—driven solely by thermal noise—to perform target computations (e.g., image classification) within a fixed observation time. We propose a gradient-descent-based parameter optimization method framed as a teacher–student paradigm: a deterministic teacher network generates ideal neural activation trajectories, while a stochastic student system models tunable thermodynamic hardware (e.g., bistable units with adjustable energy barriers and coupling strengths). Physical parameters are optimized end-to-end via backpropagation through the stochastic dynamics to minimize trajectory divergence. To our knowledge, this is the first approach to apply gradient descent directly for end-to-end training of physical thermodynamic computing substrates. Experiments demonstrate robust classification performance on MNIST, with theoretical energy consumption over seven orders of magnitude lower than conventional digital implementations. The method establishes a new paradigm for ultra-low-power, brain-inspired computing grounded in nonequilibrium thermodynamics.

This paper addresses the lack of physical interpretability in distribution shift and generalization error within machine learning. Methodologically, it systematically reconstructs thermodynamic laws within the exponential family framework, modeling log-loss minimization as a maximum-entropy-driven statistical mechanical process and establishing rigorous correspondences between thermodynamic quantities—such as work, heat, and thermodynamic cycles—and learning dynamics. Key contributions include: (i) the first formulation of universal thermodynamic laws—zeroth through fourth—for exponential families; (ii) a thermodynamic characterization of distribution shift, yielding a principled generalization error bound under shift-induced dynamics; and (iii) a computable “statistical heat engine” evaluation framework grounded in information geometry and log-loss optimization. Collectively, these results provide a novel information-physical perspective on AI foundations and introduce quantitative analytical tools for characterizing learning behavior under distributional change.

Existing building thermodynamic models rely heavily on long-term historical data and domain expertise, resulting in poor generalizability and limited transferability—hindering their applicability to real-time HVAC control. To address this, we propose a thermodynamic model integration framework specifically designed for HVAC control. Our approach introduces a novel hierarchical reinforcement learning (HRL)-based mechanism for dynamic model selection and online weighted ensemble, enabling adaptive modeling of non-stationary building time-series data. Leveraging pre-existing models as foundational components, the framework eliminates the need for de novo modeling, thereby significantly enhancing cross-building transferability and modeling efficiency. Offline evaluations and on-site deployment demonstrate that our method reduces prediction error by 32%, decreases modeling time by 76%, and achieves a 14.8% energy saving in HVAC operation.

This work addresses the lack of thermodynamic consistency validation in existing synthesizability prediction models and their reliance on scarce failed synthesis data. The authors propose a novel evaluation framework that requires no negative (failed) samples, leveraging the CHGNet interatomic potential and the Chemeleon generative model to construct a test dataset. By integrating convex hull energy and thermodynamic selectivity metrics, they systematically quantify the alignment between model predictions and fundamental thermodynamic stability and reaction selectivity principles. Their analysis reveals that mainstream synthesizability models generally overestimate the feasibility of synthesis; however, certain scoring functions exhibit trends consistent with thermodynamic indicators. These findings provide critical insights for developing physically grounded, thermodynamically consistent models for predicting material synthesizability.

This study addresses the challenge of quantifying stability and reconfiguration in neuropsychological systems by proposing a “thermoinformatics” framework inspired by statistical thermodynamics, mapping macroscopic thermodynamic variables—entropy, internal energy, temperature, and Helmholtz free energy—to neural dynamical metrics. Methodologically, it integrates multichannel synchronized EEG with behavioral data to construct an information-theoretic thermodynamic model, enabling dynamic tracking of free energy and state-space trajectory analysis. Its key contribution lies in the first systematic application of thermodynamic principles to cross-species neurobehavioral coupling studies—specifically, maternal-infant EEG during the A-not-B task and optogenetically manipulated juvenile mice—successfully decoupling neural reconfiguration from behavioral output: decision errors correlate with elevated information heat, whereas correct choices precede declines in temperature and free energy. Results demonstrate the framework’s cross-scale and cross-species generality, offering a novel theoretical foundation and computationally tractable metrics for investigating cognitive stability and adaptive neural reconfiguration.

This work addresses the inefficiency and limited scalability of traditional thermodynamic cycle design, which often relies on expert intuition or exhaustive search. To overcome these limitations, the authors propose an end-to-end automated co-design framework that represents cycles as grammar-constrained graph structures and integrates graph neural networks, physics-informed surrogate models, and hierarchical reinforcement learning with a Manager-Worker architecture to jointly optimize both topology and parameters. This approach uniquely unifies graph-based representation, physical surrogates, and reinforcement learning, offering both high efficiency and broad applicability. Validated on heat pump and heat engine benchmarks, the method not only reproduces established cycles but also discovers 18 and 21 novel configurations, respectively, achieving performance improvements of 4.6% and 133.3% over baseline designs.

This study addresses phase lag during rapid cloud transients and non-physical nighttime power predictions in off-grid photovoltaic systems by proposing a thermodynamic liquid manifold network architecture. The method embeds 15-dimensional meteorological and geometric variables into a Koopman-linearized Riemannian manifold, integrating a spectral calibration unit with a thermodynamic Alpha gating mechanism. By incorporating real-time atmospheric opacity and a theoretical clear-sky boundary model, the architecture structurally enforces celestial geometric constraints. Evaluated on five years of semi-arid climate data, the model achieves an RMSE of 18.31 Wh/m² and a Pearson correlation coefficient of 0.988, maintains zero prediction error over 1,826 nights, exhibits response delays under 30 minutes during abrupt weather changes, and contains only 63,458 parameters—demonstrating markedly improved physical consistency and dynamic response accuracy.

This work addresses the absence of a thermodynamic interpretation for existing stochastic differential equation (SDE)-based generative models—such as diffusion models and Schrödinger bridges—within the framework of nonequilibrium statistical mechanics. By extending the classical Jarzynski equality to scenarios involving time-varying temperature and non-conservative driving forces, the study introduces, for the first time, trajectory-level definitions of work, heat, and entropy production. It derives a generalized Jarzynski equality and a second-law-like inequality, thereby embedding SDE generative models into the formalism of stochastic thermodynamics. Leveraging tools from stochastic calculus, path integrals, and nonequilibrium statistical mechanics, this work establishes a comprehensive thermodynamic formulation for SDE-based generative modeling, offering deeper insight into their physical underpinnings and opening new avenues for the design and analysis of such models.

This study addresses critical limitations in deep learning–based solar irradiance forecasting for off-grid photovoltaic systems, particularly spurious nighttime generation and phase lag during abrupt cloud transitions. To overcome these issues, the authors propose a lightweight prediction framework that integrates physical priors by embedding 22-dimensional meteorological and celestial geometric variables into a Koopman-linearized Riemannian manifold. The architecture incorporates a spectral calibration unit and a thermodynamic Alpha gating mechanism to explicitly encode atmospheric opacity and clear-sky boundary constraints. This approach achieves the first explicit modeling of celestial mechanics and thermodynamic principles within a compact neural network, entirely eliminating nighttime prediction errors (zero bias over 1,826 days) and attaining sub-30-minute response latency under rapidly changing weather conditions. Evaluated over five years in a semi-arid climate, the model achieves an RMSE of 18.31 Wh/m² and a Pearson correlation coefficient of 0.988 with only 63,458 parameters.