Simeng Han, Hailey Schoelkopf, Yilun Zhao et al. · salesforce
Large language models (LLMs) have achieved remarkable performance on a variety of natural language understanding tasks. However, existing benchmarks are inadequate in measuring the complex logical reasoning capabilities of a model. We present FOLIO, a human-annotated, logically complex and diverse dataset for reasoning in natural language (NL), equipped with first-order logic (FOL) annotations. FOLIO consists of 1,430 examples (unique conclusions), each paired with one of 487 sets of premises used to deductively reason for the validity of each conclusion. The logical correctness of the premises and conclusions is ensured by their FOL annotations, which are automatically verified by an FOL inference engine. In addition to the main NL reasoning task, NL-FOL pairs in FOLIO constitute a new NL-FOL translation dataset. Our experiments on FOLIO systematically evaluate the FOL reasoning ability of supervised fine-tuning on medium-sized language models. For both NL reasoning and NL-FOL translation, we benchmark multiple state-of-the-art language models. Our results show that a subset of FOLIO presents a challenge for one of the most capable {Large Language Model (LLM)} publicly available, GPT-4.
The calculation of electromagnetic field distributions within structured media is central to the optimization and validation of photonic devices. We introduce WaveY-Net, a hybrid data- and physics-augmented convolutional neural network that can predict electromagnetic field distributions with ultra fast speeds and high accuracy for entire classes of dielectric photonic structures. This accuracy is achieved by training the neural network to learn only the magnetic near-field distributions of a system and to use a discrete formalism of Maxwell's equations in two ways: as physical constraints in the loss function and as a means to calculate the electric fields from the magnetic fields. As a model system, we construct a surrogate simulator for periodic silicon nanostructure arrays and show that the high speed simulator can be directly and effectively used in the local and global freeform optimization of metagratings. We anticipate that physics-augmented networks will serve as a viable Maxwell simulator replacement for many classes of photonic systems, transforming the way they are designed.
We present a non-convex optimization algorithm metaheuristic, based on the training of a deep generative network, which enables effective searching within continuous, ultra-high dimensional landscapes. During network training, populations of sampled local gradients are utilized within a customized loss function to evolve the network output distribution function towards one peak at high-performing optima. The deep network architecture is tailored to support progressive growth over the course of training, which allows the algorithm to manage the curse of dimensionality characteristic of high-dimensional landscapes. We apply our concept to a range of standard optimization problems with dimensions as high as one thousand and show that our method performs better with fewer function evaluations compared to state-of-the-art algorithm benchmarks. We also discuss the role of deep network over-parameterization, loss function engineering, and proper network architecture selection in optimization, and why the required batch size of sampled local gradients is independent of problem dimension. These concepts form the foundation for a new class of algorithms that utilize customizable and expressive deep generative networks to solve non-convex optimization problems.
Surrogate neural network-based partial differential equation (PDE) solvers have the potential to solve PDEs in an accelerated manner, but they are largely limited to systems featuring fixed domain sizes, geometric layouts, and boundary conditions. We propose Specialized Neural Accelerator-Powered Domain Decomposition Methods (SNAP-DDM), a DDM-based approach to PDE solving in which subdomain problems containing arbitrary boundary conditions and geometric parameters are accurately solved using an ensemble of specialized neural operators. We tailor SNAP-DDM to 2D electromagnetics and fluidic flow problems and show how innovations in network architecture and loss function engineering can produce specialized surrogate subdomain solvers with near unity accuracy. We utilize these solvers with standard DDM algorithms to accurately solve freeform electromagnetics and fluids problems featuring a wide range of domain sizes.
Innovation in nanophotonics currently relies on human experts who synergize specialized knowledge in photonics and coding with simulation and optimization algorithms, entailing design cycles that are time-consuming, computationally demanding, and frequently suboptimal. We introduce MetaChat, a multi-agentic design framework that can translate semantically described photonic design goals into high-performance, freeform device layouts in an automated, nearly real-time manner. Multi-step reasoning is enabled by our Agentic Iterative Monologue (AIM) paradigm, which coherently interfaces agents with code-based tools, other specialized agents, and human designers. Design acceleration is facilitated by Feature-wise Linear Modulation-conditioned Maxwell surrogate solvers that support the generalized evaluation of metasurface structures. We use freeform dielectric metasurfaces as a model system and demonstrate with MetaChat the design of multi-objective, multi-wavelength metasurfaces orders of magnitude faster than conventional methods. These concepts present a scientific computing blueprint for utilizing specialist design agents, surrogate solvers, and human interactions to drive multi-physics innovation and discovery.
Generative adversarial networks, which can generate metasurfaces based on a training set of high performance device layouts, have the potential to significantly reduce the computational cost of the metasurface design process. However, basic GAN architectures are unable to fully capture the detailed features of topologically complex metasurfaces, and generated devices therefore require additional computationally-expensive design refinement. In this Letter, we show that GANs can better learn spatially fine features from high-resolution training data by progressively growing its network architecture and training set. Our results indicate that with this training methodology, the best generated devices have performances that compare well with the best devices produced by gradient-based topology optimization, thereby eliminating the need for additional design refinement. We envision that this network training method can generalize to other physical systems where device performance is strongly correlated with fine geometric structuring.
Metasurfaces are subwavelength-structured artificial media that can shape and localize electromagnetic waves in unique ways. The inverse design of these devices is a non-convex optimization problem in a high dimensional space, making global optimization a major challenge. We present a new type of population-based global optimization algorithm for metasurfaces that is enabled by the training of a generative neural network. The loss function used for backpropagation depends on the generated pattern layouts, their efficiencies, and efficiency gradients, which are calculated by the adjoint variables method using forward and adjoint electromagnetic simulations. We observe that the distribution of devices generated by the network continuously shifts towards high performance design space regions over the course of optimization. Upon training completion, the best generated devices have efficiencies comparable to or exceeding the best devices designed using standard topology optimization. Our proposed global optimization algorithm can generally apply to other gradient-based optimization problems in optics, mechanics and electronics.
We present a global optimizer, based on a conditional generative neural network, which can output ensembles of highly efficient topology-optimized metasurfaces operating across a range of parameters. A key feature of the network is that it initially generates a distribution of devices that broadly samples the design space, and then shifts and refines this distribution towards favorable design space regions over the course of optimization. Training is performed by calculating the forward and adjoint electromagnetic simulations of outputted devices and using the subsequent efficiency gradients for backpropagation. With metagratings operating across a range of wavelengths and angles as a model system, we show that devices produced from the trained generative network have efficiencies comparable to or better than the best devices produced by adjoint-based topology optimization, while requiring less computational cost. Our reframing of adjoint-based optimization to the training of a generative neural network applies generally to physical systems that can utilize gradients to improve performance.
A key challenge in metasurface design is the development of algorithms that can effectively and efficiently produce high performance devices. Design methods based on iterative optimization can push the performance limits of metasurfaces, but they require extensive computational resources that limit their implementation to small numbers of microscale devices. We show that generative neural networks can train from images of periodic, topology-optimized metagratings to produce high-efficiency, topologically complex devices operating over a broad range of deflection angles and wavelengths. Further iterative optimization of these designs yields devices with enhanced robustness and efficiencies, and these devices can be utilized as additional training data for network refinement. In this manner, generative networks can be trained, with a onetime computation cost, and used as a design tool to facilitate the production of near-optimal, topologically-complex device designs. We envision that such data-driven design methodologies can apply to other physical sciences domains that require the design of functional elements operating across a wide parameter space.