Long-range interactions are essential for the correct description of complex systems in many scientific fields. The price to pay for including them in the calculations, however, is a dramatic increase in the overall computational costs. Recently, deep graph networks have been employed as efficient, data-driven models for predicting properties of complex systems represented as graphs. These models rely on a message passing strategy that should, in principle, capture long-range information without explicitly modeling the corresponding interactions. In practice, most deep graph networks cannot really model long-range dependencies due to the intrinsic limitations of (synchronous) message passing, namely oversmoothing, oversquashing, and underreaching. This work proposes a general framework that learns to mitigate these limitations: within a variational inference framework, we endow message passing architectures with the ability to adapt their depth and filter messages along the way. With theoretical and empirical arguments, we show that this strategy better captures long-range interactions, by competing with the state of the art on five node and graph prediction datasets.

Inverse design, where we seek to design input variables in order to optimize an underlying objective function, is an important problem that arises across fields such as mechanical engineering to aerospace engineering. Inverse design is typically formulated as an optimization problem, with recent works leveraging optimization across learned dynamics models. However, as models are optimized they tend to fall into adversarial modes, preventing effective sampling. We illustrate that by instead optimizing over the learned energy function captured by the diffusion model, we can avoid such adversarial examples and significantly improve design performance. We further illustrate how such a design system is compositional, enabling us to combine multiple different diffusion models representing subcomponents of our desired system to design systems with every specified component. In an N-body interaction task and a challenging 2D multi-airfoil design task, we demonstrate that by composing the learned diffusion model at test time, our method allows us to design initial states and boundary shapes that are more complex than those in the training data. Our method generalizes to more objects for N-body dataset and discovers formation flying to minimize drag in the multi-airfoil design task. Project website and code can be found at https://github.com/AI4Science-WestlakeU/cindm.

The ability to perform fast and accurate atomistic simulations is crucial for advancing the chemical sciences. By learning from high-quality data, machine-learned interatomic potentials achieve accuracy on par with ab initio and first-principles methods at a fraction of their computational cost. The success of machine-learned interatomic potentials arises from integrating inductive biases such as equivariance to group actions on an atomic system, e.g., equivariance to rotations and reflections. In particular, the field has notably advanced with the emergence of equivariant message passing. Most of these models represent an atomic system using spherical tensors, tensor products of which require complicated numerical coefficients and can be computationally demanding. Cartesian tensors offer a promising alternative, though state-of-the-art methods lack flexibility in message-passing mechanisms, restricting their architectures and expressive power. This work explores higher-rank irreducible Cartesian tensors to address these limitations. We integrate irreducible Cartesian tensor products into message-passing neural networks and prove the equivariance and traceless property of the resulting layers. Through empirical evaluations on various benchmark data sets, we consistently observe on-par or better performance than that of state-of-the-art spherical and Cartesian models.

Simulating the time evolution of Partial Differential Equations (PDEs) of large-scale systems is crucial in many scientific and engineering domains such as fluid dynamics, weather forecasting and their inverse optimization problems. However, both classical solvers and recent deep learning-based surrogate models are typically extremely computationally intensive, because of their local evolution: they need to update the state of each discretized cell at each time step during inference. Here we develop Latent Evolution of PDEs (LE-PDE), a simple, fast and scalable method to accelerate the simulation and inverse optimization of PDEs. LE-PDE learns a compact, global representation of the system and efficiently evolves it fully in the latent space with learned latent evolution models. LE-PDE achieves speedup by having a much smaller latent dimension to update during long rollout as compared to updating in the input space. We introduce new learning objectives to effectively learn such latent dynamics to ensure long-term stability. We further introduce techniques for speeding-up inverse optimization of boundary conditions for PDEs via backpropagation through time in latent space, and an annealing technique to address the non-differentiability and sparse interaction of boundary conditions. We test our method in a 1D benchmark of nonlinear PDEs, 2D Navier-Stokes flows into turbulent phase and an inverse optimization of boundary conditions in 2D Navier-Stokes flow. Compared to state-of-the-art deep learning-based surrogate models and other strong baselines, we demonstrate up to 128x reduction in the dimensions to update, and up to 15x improvement in speed, while achieving competitive accuracy.

To achieve digital transformation, it is becoming more important to flexibly and quickly provide information and communication technology (ICT) services according to various business requirements. Closed-loop control and its automation with artificial intelligence/machine learning (AI/ML) is a promising approach. However, most of such efforts focus on monitoring and analysis of resources. The challenges remain in automating operation of an entire service that consists of many resources in accordance with the results of the monitoring and analysis. We propose a scheme for automatically designing the topology and configuration of an entire service with various functional/non-functional requirements. To increase both the flexibility of the requirements and quickness to design, our scheme combines a model-based approach with AI/ML. It generates a huge number of design candidates through the model-based approach, and selects promising candidates efficiently by using AI/ML. We present the architecture of the proposed scheme and experiments we conducted involving a case study. Through the experiment, we confirmed that (1) the design speed was improved by 6 times autonomously, and (2) the increasing speed of design time with respect to the increase of intent size is significantly suppressed.

A configuration of an information technology/network (IT/NW) service is composed of components (e.g., applications, servers, and switches), relationships among them, and their attributes. To run a service appropriately, its configuration should be carefully organized to satisfy customer requirements, as well as dependencies and constraints derived from the components. Designing such a service configuration is daunting when the service is deployed in a heterogeneous environment including multiple clouds, edge devices, and different types of existing servers and network nodes. Thus, the time and cost for the designing task are serious problems in providing new services. In this paper, we present Weaver, an automated service configuration designer that generates a concrete service configuration on the basis of abstract customer requirements and the environment in which the service is deployed. Weaver accepts information about such requirements and environment as input and converts it into a fully concretized service configuration that can be deployed in the designated environment. In this paper, we take an example of a video surveillance Internet of Things (IoT) service, where multiple components including cameras, video analyzers and terminals, are distributed over multiple locations. In the evaluation, we show that Weaver produces configurations for designated heterogeneous environments. We also show that design time is dramatically shorter than in traditional system integration procedures. For example, the design time is shorter than a minute when the number of components is less than 150.

In an effort to study and characterize scattering mechanisms of debris particles in tornadoes, a numerical polarimetric radar emulator was developed. This paper is primarily motivated by attempts to explain radar observations near tornadoes. One such observation is the regions of negative differential reflectivity, which have been found near tornadoes but they are yet to be explained physically. There are hypotheses that suggest common debris alignment and/or dominant scattering from objects with high radar-cross-section (RCS) values that cause negative ZDR, but they are extremely challenging to verify due to the inherent danger near the vicinity of tornadoes. It is, however, possible to numerically construct the scenes through representative simulations to verify the plausible causes. This serves as our primary motivation to develop the radar emulator. The novel aspects of this paper are the realistic trajectory derivation, which is based on a physical air-drag model, and the representative diversity of RCS contributions from each debris object, developed through realistic polarimetric RCS modeling and anechoic chamber measurements.

Character projection (CP) is a high-speed mask-less exposure technique for electron-beam direct writing (EBDW). In CP exposure of VIA layers, higher throughput is realized if more VIAs are exposed in each EB shot, but it will result in huge number of VIA characters required for arbitrary VIA placement. We adopt one-dimensional VIA array as the basic CP character architecture to increase VIA numbers in an EB shot while saving the stencil area by superposed character arrangement. CP throughput is further improved by layout constraints for VIA placement in detail routing phase. Our experimental results give estimated EB shot counts less than 174G shot/wafer in 14nm technologies.

To improve throughput of Electron Beam Direct Writing (EBDW) with Character Projection (CP) method, we propose a structured routing architecture (SRA) where VIA placement and wire-track interchange is restricted so as to reduce possible layout patterns in VIA and metal layers. CP exposure is accelerated by the increased figure numbers of VIAs and metal segments in each CP character due to the reduced character variations. We demonstrate a design flow that enables routing on the interconnect layer with the limited flexibility using a commercial routing tool and a few own programs. We also present a methodology to design character stencils with improved area efficiency by character superposition. Our experimental results proved the architecture's feasibility in achieving the target EBDW performance in 14nm technologies.

We have developed an laser ablation system with an automatic focusing (AF) and robotic scanning mechanism for precision malignant gliomas resection in neurosurgery. A 5-aminolevulinic acid (5-ALA)-induced fluorescence based intra-operative tumor diagnosis technique has been incorporated into the robotic laser ablation system. The system enables an intra-operative identification of the position of a tumor with fluorescence illuminated by a laser excitation. The AF and robotic scanning mechanism assists in tracking the surface of the malignant brain tumors, and provides position information for both laser scanning in 5-ALA fluorescence identification and laser ablation in tumor resection. Experimental results showed the automatic focus device had a precision of 0.5 mm and the mechanism could track the brain surface with a movement caused by pulsation and respiration.

We propose a co-channel inter-cluster interference canceller for distributed antenna systems (DAS) that utilize the multiuser Multiple Input Multiple Output (MU-MIMO) technique in the downlink. We first define the cluster as a virtual cell that consists of a few remote antennas and subscriber stations which communicate cooperatively by MU-MIMO. Multiuser beamforming is applied to each cluster, so the proposed method can decrease the computational complexity of spatial division multiplexing processes compared to MIMO systems with large number of antennas. Though clustering causes co-channel inter-cluster interference, it can be cancelled simply by linear processing of the downlink transmission signals. Computer simulations show that the proposed method enables denser cluster configurations. The method contributes to the practical realization of distributed antenna systems through its effective frequency reuse.

We propose an effective resource allocation method to handle OFDMA and multiuser MIMO. This method offers two optimization techniques. In the frequency domain, those subcarriers that are useful for one subscriber station (SS) are allocated to that SS. In the spatial domain, several SSs are selected for spatial multiplexing. In frequency selective fading environments, the number of transmittable bits increases since the frequency and spatial resources can allocated efficiently. We define a permission table that indicates the allocation status. It is optimized using a genetic algorithm. The table is associated with GA chromosomes. They are renewed so as to improve two criteria, total throughput and minimum throughput among SSs. Calculations show that the proposed method improves the total throughput performance by up to around 50% compared to random resource allocation. The minimum throughput of each SS can also be improved with a small penalty in total throughput.

This work proposes a simple algorithm for controlling baseband predistortion in quasimillimeter wave fixed wireless access (FWA) systems. This algorithm selects the optimal distortion compensation data from multiple sets of distortion compensation data in order to reduce the detected out-band emission power. This algorithm has features that prevent severe degradation even though it does not use the strict distortion compensation data update method. The effectiveness of the proposed algorithm is verified by a simulation that uses 64QAM as the modulation scheme; the saturated output power of the AIM-AM characteristics is 0.5 dB, the degree of phase shift for the AM-PM characteristic is 5%, and 6 dB back off is used. The simulation results show that the degradation in characteristics based on distortion calculated as distortion compensation data error is approximately 0.