Models for spatio-temporal data prediction have wide applications in different fields, ranging from media content generation to weather forecasting. However, not all aspects of model design for video-like sequences are well-researched, particularly in long-term forecasting tasks involving datasets characterized by strong periodicity, as many existing benchmarks do not adequately address these challenges.This paper proposes a novel approach to enhance state-of-the-art video prediction methods specifically for long-term tasks involving periodic datasets. Our approach incorporates an FFT-based algorithm to assess the periodicity of video data, which is then utilized to estimate the optimal number of input features for the model. We validate it across two types of deep learning architectures for video prediction—Convolutional Neural Networks (CNNs) and diffusion models—using benchmark datasets that encompass both media data (Moving MNIST, KTH) and metocean data (WeatherBench, OSISAF).Our experimental results demonstrate that the proposed approach leads to substantial improvements in prediction quality, achieving over a 30% reduction in Mean Absolute Error (MAE) and more than a 15% increase in Peak Signal-to-Noise Ratio (PSNR) for the models considered. It confirms the effectiveness of periodicity analysis in enhancing video prediction performance, paving the way for more accurate long-term forecasting in various applications. The source code is available on GitHub - https://github.com/ITMO-NSS-team/IJCNN2025_periodicity.

Predictive modeling of sea ice conditions in the Arctic region is important task for environmental monitoring, climate change issue and offshore oil production. The existing physics-based and data-driven solutions for ice forecasting are usually not flexible enough to satisfy the domain-specific requirements. Therefore, we propose the lightweight adaptive modeling approach named LANE-SI (Lightweight Automated Neural Ensembling for Sea Ice). It use ensemble of simple encoder-decoder architecture deep learning models with differ-ent loss functions for forecasting of spatial distribution for sea ice concentration in the specified water area. Experimental studies confirm the quality of a long-term fore-cast based on a deep learning model fitted to the specific water area is comparable to resource-intensive physical modeling, and for some periods of the year, it is superior. We achieved a 20 % improvement against the state-of-the-art physics-based forecast system SEAS5 for the Kara Sea.

The effective modelling for multi-modal data is still one of challenging problems in modern machine learning. The most existing tools for automated machine learning avoids the multi-modality, and a few others use computationally expensive end-to-end deep models, that can be redunant for a lot of cases. In this paper, we propose the evolutionary approach for automated design of light-weight multi-modal pipelines. It is based an graph-based representation of pipeline and simplification of over-complicated solutions. The experiments with AutoML Multi-modal Benchmark confirms that the proposed approach allow achieving the competitive performance against more complicated state-of-the-art solutions.

Co-crystallization is an accessible way to control physicochemical characteristics of organic crystals, which finds many biomedical applications. In this work, we present Generative Method for Co-crystal Design (GEMCODE), a novel pipeline for automated co-crystal screening based on the hybridization of deep generative models and evolutionary optimization for broader exploration of the target chemical space. GEMCODE enables fast *de novo* co-crystal design with target tabletability profiles, which is crucial for the development of pharmaceuticals. With a series of experimental studies highlighting validation and discovery cases, we show that GEMCODE is effective even under realistic computational constraints. Furthermore, we explore the potential of language models in generating co-crystals. Finally, we present numerous previously unknown co-crystals predicted by GEMCODE and discuss its potential in accelerating drug development.

Resource-intensive computations are a major factor that limits the effectiveness of automated machine learning solutions. In the paper, we propose a modular approach that can be used to increase the quality of evolutionary optimization for modelling pipelines with a graph-based structure. It consists of several stages - parallelization, caching, and evaluation. Heterogeneous and remote resources can be involved in the evaluation stage. The conducted experiments confirm the correctness and effectiveness of the proposed approach. The implemented algorithms are available as a part of the open-source framework FEDOT.

In the paper, we discuss the applicability of automated machine learning for the effective multi-scale modeling of the industrial sensors time series. The proposed approach is based on the evolutionary generative design of the composite modeling pipelines. The iterative data decomposition algorithm is proposed in the paper to improve the quality of the sensor time series forecasting. To effectively use it in an automated way, the boosting-like mutation operators have been implemented for graphs-based genotypes. The proposed approach reduced the forecast error by 10% compared to the competitor library AutoTS. Also, the proposed modifications of the evolutionary algorithm resulted in better metrics in 78% of the cases where they were used.

In this paper, a multi-objective evolutionary surrogate-assisted approach for the fast and effective generative design of coastal breakwaters is proposed. To approximate the computationally expensive objective functions, the deep convo-lutional neural network is used as a surrogate model. This model allows optimizing a configuration of breakwaters with a different number of structures and segments. In addition to the surrogate, an assistant model was developed to estimate the confidence of predictions. The proposed approach was tested on the synthetic water area, the SWAN model was used to calculate the wave heights. The experimental results confirm that the proposed approach allows to obtain more effective (less expensive with better protective properties) solutions than non-surrogate approaches for the same time.

In this paper, a multi-objective approach for the design of composite data-driven mathematical models is proposed. It allows automating the identification of graph-based heterogeneous pipelines that consist of different blocks: machine learning models, data preprocessing blocks, etc. The implemented approach is based on a parameter-free genetic algorithm (GA) for model design called GPComp@Free. It is developed to be part of automated machine learning solutions and to increase the efficiency of the modeling pipeline automation. A set of experiments was conducted to verify the correctness and efficiency of the proposed approach and substantiate the selected solutions. The experimental results confirm that a multi-objective approach to the model design allows us to achieve better diversity and quality of obtained models. The implemented approach is available as a part of the open-source AutoML framework FEDOT.