1.1.

Physics and AI

The interaction between physics and AI has been a symbiotic one, where principles from physics have been applied to enhance AI models, and AI, in turn, has been used to solve complex problems in physics. This dynamic interplay is beautifully exemplified by the work of the 2024 Nobel laureates in Physics, John Hopfield³ and Geoffrey Hinton.⁴ Some of their pioneering contributions to artificial neural networks,^5–16 which are deeply rooted in concepts borrowed from or inspired by physics,^6,7,11,12 have laid the foundations for the modern AI revolution; see Fig. 1. John Hopfield, a physicist by training, was one of the first to draw a strong connection between physics and neural networks. His work was inspired by the complex world of spin glasses—disordered magnetic systems with intricate interactions. Hopfield recognized an analogy between these physical systems and networks of interconnected neurons in the brain.⁷ This insight led him to develop a type of recurrent neural network, known as the Hopfield network,¹⁷ capable of storing and retrieving patterns [Fig. 1(a)]. His 1982 paper,⁷ a cornerstone in the field of AI, demonstrated how principles from condensed matter physics could be harnessed to create computational systems that have learning and memory. The Hopfield network provided methods to explore how associative memory¹⁸ works, both in biological systems and artificial ones, and became one of the foundational models for AI research. Hopfield’s work was groundbreaking not just because it connected physics to AI but because it introduced the concept of energy landscapes^9,19 to neural networks. In a Hopfield network, the system settles into states of minimum energy, akin to how a physical system seeks equilibrium. This analogy to physical systems allowed researchers to use well-established methods from statistical mechanics to analyze and better understand neural networks, which opened up new avenues for creating advanced AI systems that could more closely emulate human cognition.

Fig. 1

Some of the major contributions of John Hopfield and Geoffrey Hinton, with their connections to physical systems. (a) Left: the Hopfield model, a recurrent neural network capable of storing and retrieving patterns. John Hopfield proposed an energy-based memory model capable of storing and retrieving patterns, drawing intuition from the energy dynamics of spin glass. The network’s dynamics are driven by minimizing the energy of the model state $E(s)$. Center: the energy landscape of a Hopfield network, depicting how the network converges to stable states (attractors) represented as valleys in the energy landscape. Right: a spin glass system, i.e., a disordered material with magnetic interactions among atomic spins. The energy of a state in the Hopfield model is similar to the energy Hamiltonian $H(\sigma )$ of the spin glass state $\sigma$. (b) Left: the Boltzmann machine, a stochastic generalization of the Hopfield network, includes hidden units (represented by gray nodes) that enhance its representation capability. By incorporating the Boltzmann distribution from statistical mechanics, Geoffrey Hinton introduced stochasticity into his neural network models, enhancing their ability to learn complex patterns. Right: the Boltzmann distribution, which governs the probability of a state based on its energy. This distribution plays a key role in the stochastic activation of artificial neurons in Boltzmann machines.

Similar to Hopfield’s scientific explorations at the intersection of physics and AI, Geoffrey Hinton, a cognitive psychologist and computer scientist, also took the push–pull relationship between AI and physics as a major inspiration for his seminal work. Hinton recognized the potential of Boltzmann machines,¹² a type of stochastic neural network inspired by statistical mechanics, to learn complex patterns from data [Fig. 1(b)]. One of Hinton’s groundbreaking contributions was developing efficient learning algorithms for these networks.^14,15,20 These algorithms enabled neural networks to extract meaningful features from data, such as images, text, or language, by optimizing the network’s parameters through processes similar to energy minimization. He also popularized the backpropagation algorithm,^13,21,22 revolutionized convolutional neural networks,²³ and introduced techniques such as dropout to improve training.²⁴ Hinton’s work laid the groundwork for modern deep learning architectures, the applications of which have revolutionized fields such as computer vision, natural language processing, robotics, and biomedical sciences.

Beyond the seminal works of Hopfield and Hinton, the influence of physics on AI extends to various other areas. For instance, the concept of renormalization group, a powerful tool in condensed-matter and particle physics used to study systems with many interacting components across different scales, has found applications in deep learning for analyzing hierarchical structures and improving the efficiency of training algorithms.²⁵ Another example is the use of quantum mechanics principles to develop new types of neural networks, known as quantum neural networks,²⁶ which leverage quantum phenomena such as superposition and entanglement to potentially achieve exponential speedups for certain computational tasks. Furthermore, ideas from information theory, a field with deep roots in thermodynamics and statistical mechanics, have been instrumental in developing algorithms for compressing and efficiently representing information in AI systems. As another important example, diffusion models,²⁷ a new class of powerful generative models, draw direct inspiration from the physics of diffusion^28,29 and Brownian motion. These examples illustrate the rich and ongoing cross-fertilization between physics and AI, where fundamental concepts from physics continue to inspire novel approaches and solutions in the realm of AI.

2.1.

AI in Computational Imaging and Sensing

One of the exciting areas where AI has been making a significant impact is computational imaging and sensing. Traditional imaging methods often face limitations due to the physical constraints of optics, such as resolution limits or noise. AI, however, is offering powerful new tools to mitigate some of these barriers. In microscopy, for example, AI algorithms can enhance image resolution, remove noise and artifacts, and even reconstruct 3D structures from limited data.³³ Techniques such as super-resolution microscopy, which breaks the diffraction limit of light to reveal finer details than previously possible, have been significantly advanced by AI.^34–40 In holographic imaging, AI algorithms have excelled in solving complex physics-based inverse problems,^41–46 such as reconstructing a 3D scene from holographic data,⁴⁷ with greater accuracy and speed than traditional methods, also providing different contrast mechanisms, e.g., reconstructing the images of specimens with brightfield contrast using their monochrome holograms.⁴⁷ In fact, AI has been driving major innovations through such cross-modality image transformations,^36,48 where the spatial and spectral information typically associated with one imaging modality is extracted from data acquired using a different modality. This capability is opening up exciting new possibilities for biomedical imaging⁴⁹ and remote sensing,⁵⁰ among others. A compelling example is virtual staining in digital pathology and microscopy.^51,52 Traditional histological staining of tissue involves applying chemical stains to biological tissue to highlight various features under a microscope. However, this staining process can be time-consuming, laborious, and costly and can also damage the samples. Deep neural networks can now routinely transform label-free images of specimens into virtually stained microscopic images that mimic the appearance of traditionally stained images, eliminating the need for the chemical staining processes.^49,53 This allows for faster, cheaper, and more efficient analysis of biological samples and has significant implications for histology as well as live-cell imaging, where minimizing or eliminating chemical perturbations to the native biological system (through, e.g., external labeling and tags) is crucial.

AI is also making significant inroads in optical sensing, impacting both the design of sensors and the interpretation of sensor data.⁵⁴ In areas such as biosensing and environmental monitoring, AI algorithms can rapidly process complex optical signals to detect subtle changes and identify specific analytes or conditions with greater sensitivity and specificity.^55–63 AI is also being used to design novel optical sensors with improved performance. For instance, in the development of optical sensors for point-of-care diagnostics, AI can optimize the design of the optical detection systems to enhance sensitivity/specificity and reduce sample volume requirements, while also providing multiplexed detection that can be used for the rapid and quantitative measurement of a panel of biomarkers and disease conditions.^63–68 By automating the optimization, quantitative multiplexed sensing, and decision processes, AI is accelerating the development of innovative optical sensors with tailored functionalities for various applications in point-of-care sensing, diagnostics, and environmental monitoring, as well as structural health, among many others.^54,69–72

2.2.

AI-Driven Optics and Photonics Design

AI is also revolutionizing the design of optical materials, devices, and systems^73–76 by enabling a paradigm shift in “inverse” design.^30,77 Traditional inverse design approaches in optics and photonics typically rely on iterative optimization algorithms. These algorithms start with an initial guess of the device structure and repeatedly simulate its performance, using the results to refine the design parameters. This process continues until the desired performance metric is achieved. While these methods can be effective, they often require significant computational resources and time, especially for complex tasks and designs. In contrast, deep learning-based approaches offer more efficient and powerful alternatives. These alternative methods employ training a neural network on a large dataset of optical structures and their corresponding performance metrics.⁷⁸ Once trained, the network can rapidly predict the performance of new designs and even generate novel structures with desired properties. This learning-based approach significantly accelerates the design process and enables the exploration/optimization of a wider range of parameters and possibilities.^79–81 AI-powered inverse design has already led to the creation of materials and systems with unprecedented capabilities. These include unidirectional imagers,⁸² invisibility cloaks that can render objects invisible,⁸³ and ultra-efficient light absorbers for enhanced solar energy harvesting,⁸⁴ among many others. Furthermore, this AI-powered optimization framework allows for the smart design of free-form optics, enabling the creation of compact and lightweight optical systems with superior performance.^85–87