SciRS2 Scientific Computing and AI/ML Infrastructure

The scirs2-graph library is designed for a wide range of graph processing tasks, spanning from academic research to production-scale workloads. Key application areas include:

Core Applications

• Social Network Analysis: Implementing centrality measures and community detection.
• Route Finding: Calculating shortest paths, A* search, and alternative routes.
• Machine Learning: Extracting graph features and generating graph embeddings.
• Bioinformatics: Analyzing protein networks and functional modules.
• Operations Research: Solving network flow and supply chain optimization problems.
• Visualization: Utilizing layout algorithms and exporting graph data.

Advanced Scenarios

• Temporal Graphs: Analyzing time-respecting paths and temporal motifs.
• Hypergraphs: Performing document analysis and identifying overlapping communities.
• Graph Comparison: Using edit distance and kernel methods.
• Multi-layer Networks: Conducting cross-platform analysis.
• Advanced Pathfinding: Implementing constrained and multi-objective optimization.

Production Patterns

• Database Integration: Efficiently loading graph data from SQL databases.
• Pipeline Processing: Building configurable analysis workflows.
• Monitoring & Logging: Integrating performance tracking and observability.
• Configuration Management: Managing flexible algorithm parameters.
• Testing & Validation: Using property-based and integration testing.
• Deployment Patterns: Implementing caching, cancellation, and health checks.

The scirs2-ndimage crate provides N-dimensional image processing tools modeled after SciPy's ndimage module. Key functional areas include:

• Filtering: Gaussian, median, rank, bilateral, edge detection (Sobel, Prewitt, etc.), and custom generic filters.
• Morphology: Binary (erosion, dilation, opening, closing) and grayscale operations, distance transforms, and connected component labeling.
• Measurements: Region statistics (mean, variance), moments (Hu moments), and geometric properties (area, centroid, bounding box).
• Segmentation: Thresholding (Otsu, adaptive), watershed, active contours, and level set methods.
• Feature Detection: Canny edges, Harris/FAST corners, SIFT descriptors, and HOG.
• Geometric Interpolation: Coordinate mapping, affine transforms, rotation, zooming, and shifting.
• Specialized Analysis: 3D volumetric analysis, medical imaging (Frangi vesselness, bone enhancement), and hyperspectral processing.

scirs2-series is a comprehensive time series analysis library for Rust, part of the SciRS2 scientific computing ecosystem. It provides a unified, type-safe API for a wide spectrum of time series tasks, ranging from classical econometrics to deep learning.

Key capabilities include:

• Forecasting: Classical models (ARIMA, ETS), Neural architectures (TFT, N-BEATS, N-HiTS, DeepAR), and streaming/online algorithms.
• Modeling: State-space models, Kalman filtering, Volatility/GARCH models, Long-memory processes, and Vector Autoregressive (VAR) models.
• Advanced Analysis: Causality, Cointegration, Functional Data Analysis (FDA), and hierarchical reconciliation.
• Uncertainty Quantification: Confidence/prediction intervals for classical models, probabilistic outputs for neural models (DeepAR), and coverage/Winkler-score diagnostics.

Design principles:

• Ecosystem Coherence: Built on scirs2-core abstractions with no C or Fortran dependencies.
• Performance: Leverages SIMD-accelerated operations via the scirs2-core simd feature. It avoids rayon dependencies, instead using std::thread directly within the out_of_core module for its worker pool.

SciRS2 is a modular ecosystem organized into several functional areas. You can use the main scirs2 crate for all-in-one access or import specific crates to reduce overhead and dependencies.

Scientific Computing Core

• scirs2-linalg: Linear algebra (Matrix operations, solvers, tensor decompositions).
• scirs2-stats: Statistics (Distributions, Bayesian methods, Gaussian processes).
• scirs2-optimize: Optimization (MIP/SDP/SOCP, Bayesian, stochastic, and convex optimization).
• scirs2-integrate: Numerical integration (ODE/BVP/DAE, PDE, SDE solvers).
• scirs2-interpolate: Interpolation (RBF, B-spline, Kriging).
• scirs2-special: Special mathematical functions (Mathieu, Jacobi theta, etc.).
• scirs2-signal: Signal processing (Kalman filtering, compressed sensing, adaptive filtering).
• scirs2-sparse: Sparse matrix operations (Eigensolvers, AMG, block preconditioners).
• scirs2-spatial: Spatial algorithms (R*-Tree, Voronoi, geodata projections).

Advanced Modules

• scirs2-cluster: Clustering (GMM, HDBSCAN, topological clustering).
• scirs2-ndimage: N-dimensional image processing (Segmentation, optical flow).
• scirs2-fft: FFT and spectral methods (Sparse FFT, Wavelets, DCT/DST).
• scirs2-io: I/O utilities (Protobuf, Parquet, HDF5-lite, Arrow).
• scirs2-datasets: Sample datasets and loaders.
• scirs2-validation: Distribution validation utilities.

AI and Machine Learning

• scirs2-autograd: Automatic differentiation (Gradients, JVP/VJP, Hessian).
• scirs2-neural: Neural networks (Transformers, GNNs, Diffusion, RL).
• scirs2-graph: Graph processing (Community detection, Max flow, Embeddings).
• scirs2-transform: Data transformation (UMAP, t-SNE, Optimal transport).
• scirs2-metrics: ML evaluation metrics (IoU, mAP, FID, fairness).
• scirs2-text: Text processing (Tokenizers, NER, Topic modeling).
• scirs2-vision: Computer vision (SLAM, 3D reconstruction, Segmentation).
• scirs2-series: Time series analysis (Deep learning forecasting, Volatility models).

scirs2-optimize is a production-ready, pure-Rust optimization library that provides a wide range of classical numerical methods and state-of-the-art algorithms. It is designed for scientific computing tasks such as model fitting, engineering design, neural network training, and logistics scheduling.

Key categories of optimization supported include:

• Continuous: Unconstrained, constrained (equality, inequality, bounds), and conic optimization.
• Discrete & Combinatorial: Mixed-integer programming (MIP), branch-and-bound, and dynamic programming.
• Global: Bayesian optimization, DIRECT, differential evolution, and metaheuristics.
• Multi-objective: NSGA-II, NSGA-III, and scalarization methods.
• Stochastic & Online: SGD variants (Adam, AdamW, etc.) and variance reduction (SVRG, SARAH).
• Structured: Game theory, bilevel optimization, minimax, and robust optimization.

The scirs2-neural crate provides a comprehensive suite of tools for deep learning in Rust, ranging from basic layers to advanced research architectures.

Core Components

• Layers & Activations: Dense/Linear, Convolutional (1D, 2D, 3D), Recurrent (LSTM, GRU), Normalization (Batch/Layer Norm), and Dropout.
• Architectures: Transformers (BERT, GPT, ViT), CNNs (ResNet, EfficientNet, ConvNeXt, MobileNet), and Seq2Seq models.
• Training & Optimization: Automatic differentiation, advanced optimizers (Adam, AdamW), learning rate scheduling, gradient clipping, and mixed precision training.
• Specialized Applications: Computer Vision (Object Detection, Segmentation), NLP (Text Classification, Sentiment Analysis), and Generative Models (VAE, GAN).

Advanced Features

• Hardware Acceleration: SIMD optimizations and CUDA-based GPU acceleration.
• Production Readiness: Model serialization/deserialization, ONNX export, and quantization.
• Visualization: Real-time training progress, model architecture diagrams, attention maps, and confusion matrices.

The scirs2-datasets module is organized into several specialized learning areas:

• Getting Started: Fundamentals of loading toy datasets (Iris, Boston, Digits), exploring dataset structures, and error handling.
• Data Generation: Creating synthetic datasets including classification, regression, clustering (blobs, hierarchical), non-linear patterns (spirals, moons, swiss roll), and time series with trends/seasonality.
• Cross-Validation: Implementing K-Fold, Stratified K-Fold, Time-series (temporal order aware), Group-based, Leave-One-Out, and Nested cross-validation.
• Dataset Utilities: Preprocessing pipelines including scaling (Standard, Min-Max, Robust), imputation, feature engineering, and sampling (Random, Stratified, Bootstrap, SMOTE).
• Custom Datasets: Loading external data from CSV, JSON, ARFF, and LIBSVM, as well as database integration (SQL/NoSQL) and streaming for large datasets.
• Performance Optimization: Advanced techniques including SIMD, GPU acceleration (CUDA), parallel processing, and memory-efficient streaming.