Journal of Computational Science最新文献

Automated code generation allows for a separation between the development of a model, expressed via a domain specific language, and lower level implementation details. Algorithmic differentiation can be applied symbolically at the level of the domain specific language, and the code generator reused to implement code required for an adjoint calculation. However the adjoint calculations are complicated by the well-known problem of storing or recomputing the forward data required by the adjoint, and different checkpointing strategies have been developed to tackle this problem. This article considers the combination of high-level algorithmic differentiation with step-based checkpointing schedules, with the primary application being for solvers of time-dependent partial differential equations. The focus is on algorithmic differentiation using a dynamically constructed record of forward operations, where the precise structure of the original forward calculation is unknown ahead-of-time. In addition, high-level approaches provide a simplified view of the model itself. This allows data required to restart and advance the forward, and data required to advance the adjoint, to be identified. The difference between the two types of data is here leveraged to implement checkpointing strategies with improved performance.

In the realm of fluid mechanics, where computationally-intensive simulations demand significant time investments, especially in predicting aerodynamic coefficients, the conventional use of time series forecasting techniques becomes paramount. Existing methods prove effective with periodic time series, yet the challenge escalates when faced with aperiodic or chaotic system responses. To address this challenge, we introduce DARSI (Deep Auto-Regressive Time Series Inference), an advanced architecture and an efficient hybrid of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) components. Evaluated against established architectures (CNN, DLinear, LSTM, LSTNet, and PatchTST) for forecasting Coefficient of Lift ($C_L$) values corresponding to Angles of Attack (AoAs) across periodic, aperiodic, and chaotic regimes, DARSI demonstrates remarkable performance, showing an average increase of 79.95% in CORR, 76.57% reduction in MAPE, 94.70% reduction in MSE, 76.18% reduction in QL, and 75.21% reduction in RRSE. Particularly adept at predicting chaotic aerodynamic coefficients, DARSI emerges as the best in static scenarios, surpassing DLinear and providing heightened reliability. In dynamic scenarios, DLinear takes the lead, with DARSI securing the second position alongside PatchTST. Furthermore, static AoAs at 24.7 are identified as the most chaotic, surpassing those at 24.9 and the study reveals a potential inflection point at AoA 24.7 in static scenarios for both DLinear and DARSI, warranting further confirmation. This research positions DARSI as an adept alternative to simulations, offering computational efficiency with significant implications for diverse time series forecasting applications across industries, particularly in advancing aerodynamic predictions in chaotic scenarios.

English language education is undergoing a transformative shift, propelled by advancements in technology. This research explores the integration of the Internet of Things (IoT) and Generative Artificial Intelligence (Generative AI) in the context of English language education, with a focus on developing a personalized oral assessment method. The proposed method leverages real-time data collection from IoT devices and Generative AI's language generation capabilities to create a dynamic and adaptive learning environment. The study addresses historical challenges in traditional teaching methodologies, emphasizing the need for AI approaches. The research objectives encompass a comprehensive exploration of the historical context, challenges, and existing technological interventions in English language education. A novel, technology-driven oral assessment method is designed, implemented, and rigorously evaluated using datasets such as Librispeech and L2Arctic. The ablation study investigates the impact of training dataset proportions and model learning rates on the method's performance. Results from the study highlight the importance of maintaining a balance in dataset proportions, selecting an optimal learning rate, and considering model depth in achieving optimal performance.

Nowadays, Machine Learning and Deep Learning methods have become the state-of-the-art approach to solve data classification tasks. In order to use those methods, it is necessary to acquire and label a considerable amount of data; however, this is not straightforward in some fields, since data annotation is time consuming and might require expert knowledge. This challenge can be tackled by means of semi-supervised learning methods that take advantage of both labelled and unlabelled data. In this work, we present new semi-supervised learning methods based on techniques from Topological Data Analysis (TDA). In particular, we have created two semi-supervised learning methods following two topological approaches. In the former, we have used a homological approach that consists in studying the persistence diagrams associated with the data using the bottleneck and Wasserstein distances. In the latter, we have considered the connectivity of the data. In addition, we have carried out a thorough analysis of the developed methods using 9 tabular datasets with low and high dimensionality. The results show that the developed semi-supervised methods outperform the results obtained with models trained with only manually labelled data, and are an alternative to other classical semi-supervised learning algorithms.

Constructing digital twins for predictive tumor treatment response models can have a high computational demand that presents a practical barrier for their clinical adoption. In this work, we demonstrate that proper orthogonal decomposition, by which a low-dimensional representation of the full model is constructed, can be used to dramatically reduce the computational time required to calibrate a partial differential equation model to magnetic resonance imaging (MRI) data for rapid predictions of tumor growth and response to chemotherapy. In the proposed formulation, the reduction basis is based on each patient’s own MRI data and controls the overall size of the “reduced order model”. Using the full model as the reference, we validate that the reduced order mathematical model can accurately predict response in 50 triple negative breast cancer patients receiving standard of care neoadjuvant chemotherapy. The concordance correlation coefficient between the full and reduced order models was 0.986 ± 0.012 (mean ± standard deviation) for predicting changes in both tumor volume and cellularity across the entire model family, with a corresponding median local error (inter-quartile range) of 4.36 % (1.22 %, 15.04 %). The total time to estimate parameters and to predict response dramatically improves with the reduced framework. Specifically, the reduced order model accelerates our calibration by a factor of (mean ± standard deviation) 378.4 ± 279.8 when compared to the full order model for a non-mechanically coupled model. This enormous reduction in computational time can directly help realize the practical construction of digital twins when the access to computational resources is limited.

The amalgamation of machine learning algorithms (ML) with computational fluid dynamics (CFD) represents a promising frontier for the advancement of fluid dynamics research. However, the practical integration of CFD with ML algorithms frequently faces challenges related to data transfer and computational efficiency. While CFD programs are conventionally scripted in Fortran or C/C++, the prevalence of Python in the machine learning domain complicates their seamless integration. To tackle these obstacles, this paper proposes a comprehensive solution. Our devised framework primarily leverages Python modules CFFI and dynamic linking library technology to seamlessly integrate ML algorithms with CFD programs, facilitating efficient data interchange between them. Distinguished by its simplicity, efficiency, flexibility, and scalability, our framework is adaptable across various CFD programs, scalable to multi-node parallelism, and compatible with heterogeneous computing systems. In this paper, we showcase a spectrum of CFD+ML algorithms based on this framework, including stability analysis of ML Reynolds stress models, bidirectional coupling between ML turbulence models and CFD programs, and online dimension reduction optimization techniques tailored for resolving unstable steady flow solutions. In addition, our framework has been successfully tested on supercomputer clusters, demonstrating its compatibility with distributed computing architectures and its ability to leverage heterogeneous computing resources for efficient computational tasks.

Multi-layer complex networks (MLCN) appears in various domains, such as, transportation, supply chains, etc. Failures in MLCN can lead to major disruptions in systems. Several research have focussed on different kinds of failures, such as, cascades, their reasons and ways to avoid them. This paper considers failures in a specific type of MLCN where the lower layer provides services to the higher layer without cross layer interaction, typical of a computer network. A three layer MLCN is constructed with the same set of nodes where each layer has different characteristics, the bottom most layer is Erdos–Renyi (ER) random graph with shortest path hop count among the nodes as gaussian, the middle layer is ER graph with higher number of edges from the previous, and the top most layer is preferential attachment graph with even higher number of edges. Both edge and node failures are considered. Failures happen with decreasing order of centralities of edges and nodes in static batch mode and when the centralities change dynamically with progressive failures. Emergent pattern of three key parameters, namely, average shortest path length (ASPL), total shortest path count (TSPC) and total number of edges (TNE) for all the three layers after node or edge failures are studied. Extensive simulations show that all but one parameters show definite degrading patterns. Surprising, ASPL for the middle layer starts showing a chaotic behaviour beyond a certain point for all types of failures.

In network function virtualization (NFV), Service Function Chaining (SFC) provides an ordered sequence of virtual network functions (VNFs) and subsequent steering of traffic flows through them to cater to end-to-end services. This paper addresses the NP-hard problem of minimum cost SFC deployment to support customer services that access the carrier network’s NFV infrastructure (NFVI) through some edge routers. To determine the mappings of VNFs to physical servers, a challenging aspect would be the inter-server latencies that may fluctuate over time because of the sharing nature of cloud data centers. To construct the SFC, we come up with three different formulations, each corresponding to a different informational assumption about the link latencies: First, a centralized integer linear programming (ILP) formulation is given under the assumption of the non-causal availability of exact and instantaneous inter-server latencies. The solution to this ILP can serve as a lower bound to benchmark more scalable and realistic schemes. Next, we give a distributed game-theoretic formulation (with service broker agents as players) which only requires the statistical knowledge of link latency fluctuations. The game provably admits a pure Nash equilibrium (NE) and can be solved iteratively through the well-known best response dynamics (BRD) algorithm. Our main novelty lies in the third formulation in which each service broker has neither instantaneous nor statistical knowledge of the latencies. Instead, it relies on a game-theoretic learning algorithm to compose its VNF chain only based on its own history of adopted decisions and experienced delays on each logical link. We prove that the proposed learning algorithm asymptotically converges to NE and evaluate its performance through simulations in terms of convergence and the impact of network parameters.

In healthcare, a remarkable progress in machine learning has given rise to a diverse range of predictive and decision-making medical models, significantly enhancing treatment efficacy and overall quality of care. These models often rely on electronic health records (EHRs) as fundamental data sources. The effectiveness of these models is contingent on the quality of the EHRs, typically presented as unstructured text. Unfortunately, these records frequently contain spelling errors, diminishing the quality of intelligent systems relying on them. In this research, we propose a method and a tool for correcting spelling errors in Russian medical texts. Our approach combines the Symmetrical Deletion algorithm with a finely tuned BERT model to efficiently correct spelling errors, thereby enhancing the quality of the original medical texts at a minimal cost. In addition, we introduce several fine-tuned BERT models for Russian anamneses. Through rigorous evaluation and comparison with existing spelling error correction tools for the Russian language, we demonstrate that our approach and tool surpass existing open-source alternatives by 7% in correcting spelling errors in sample Russian medical texts and significantly superior in automatically correcting real-world anamneses. However, the new approach is far inferior to proprietary services such as Yandex Speller and GPT-4. The proposed tool and its source code are available on GitHub ¹ and pip ² repositories. This paper is an extended version of the work presented at ICCS 2023 (Pogrebnoi et al. 2023)