The s-X-AIPI project has concluded
Author: Vassia Lazaraki, Nikos Makris, Clio Drimala
24th June 2025
The s-X-AIPI project aspired to transform the EU process and manufacturing industries by developing an innovative, open-source toolset of trustworthy self-X AI technologies. These AI systems are designed to operate with minimal human intervention, continuously self-improve to boost agility, resilience and sustainability throughout the product and process lifecycle.
s-X-AIPI supports industrial workers with smarter, faster decision-making, while promotes integration into a circular manufacturing economy. Utilising innovative AI tools enhances design, development, operation and monitoring of plants, products and value chains.
s-X-AIPI demonstrated in four industrial sectors: Asphalt, Steel, Aluminum, and Pharmaceuticals, showcasing a portfolio of trustworthy AI technologies like datasets, AI models, applications, which integrated into an open source toolset. Key components of this toolset include an AI data pipeline with automatic computing capabilities, an autonomic manager based on MAPE-K models, that supports Human In The Loop, as well as several AI systems based on continuous self-optimiσation, self-configuration, self-healing and self-protection.
Launched in May 2022 with 14 partners from 6 different conuntries, the 3-year project concluded in April 2025. An online Final Review will be conducted in the summer of 2025, highlighting its significant achievents during the project’s lifespan.
ADAPT AI-Powered Anomaly Detection – Steel Use Case
CORE IC developed ADAPT – Active Detection and Anomaly Processing with smart Thresholds - a cutting-edge anomaly detection system designed to enhance operational oversight in steel manufacturing, through state-of-the-art machine learning and automation.
Powered by a Conditional Variational Autoencoder (cVAE), ADAPT continuously monitors process data—including scrap input, in-process metrics, and product compositions—to identify deviations and present them to process experts.
The base model was trained in an unsupervised manner on large volumes of historical, unlabeled production data, an ideal approach for industrial environments where labeled anomalies are rare and difficult to obtain. A Bayesian optimisation framework was integrated into the training pipeline, to allow data-driven hyperparameter tuning. Finally, an explainability module was introduced to the inference for transparency and user trust: for each detected anomaly, ADAPT highlights the top contributing features, enabling experts to quickly understand root causes and take informed corrective action.
ADAPT’s strength lies in its ability to continuously evolve. It supports two complementary mechanisms for refining the base model:
semi-supervised learning loop, which incorporates expert feedback on the identified anomalies, through a Human-in-the-Loop (HITL) workflow. ADAPT’s active learning framework minimises the need for constant human intervention, while still leveraging expert input where it adds the most value—driving continuous operational decision support.
unsupervised adaptation strategy on new, unseen data that addresses both data drift and concept drift over time.
These fine-tuning processes, along with performance monitoring, model redeployment, and user feedback tracking, are fully automated within ADAPT’s MLOps pipeline. The system also logs and manages historical anomaly data, for tracking patterns, comparing model behavior over time, and supporting process audits or improvement initiatives. Experiment tracking, version control, and metadata management ensure that every model iteration is traceable, reproducible, and aligned with production requirements.
ADAPT End-to-end pipeline for robust, adaptive Anomaly Detection
Data Augmentation Techniques – Asphalt Use case
Software developments in the Asphalt Use Case (UC) faced a key limitation: the scarcity of high-quality laboratory test data. Although the data spans long operational periods, the overall volume remains low, making it difficult to train effective AI models for analysing asphalt behaviour and predicting performance outcomes. This lack of data particularly impacts supervised learning approaches, leading to class imbalance, reduced model robustness, and constrained generalisation. To address this challenge CORE IC developed a three-stage data augmentation pipeline, illustrated below, that expanded the original dataset (~500 rows) to approximately 103.
Three stages of our Data Augmentation Technique
The first stage focused on imputing missing values using the K-Nearest Neighbors (KNN) algorithm, selecting the five closest data points to estimate missing entries. In the second stage, Gaussian noise was added to the dataset to introduce variability and promote model robustness, while preserving the data's underlying structure. The final stage involved experimentation with three generative AI models—Variational Autoencoder (VAE), Denoising Autoencoder (DAE), and ReaLTabFormer—each used to generate synthetic records that enriched the dataset. These enhanced datasets were then evaluated for their impact on predictive performance. Among them, the dataset generated by the VAE method emerged as the most effective, significantly enhancing the model’s accuracy and predictive performance.
Dissemination, Communication and Exploitation Activities
We are just a few days away since the project wrapped up and cannot omit reflecting back on some valuable dissemination and communication (D&C) achievements. Over the past 3-years, CORE IC devised and led the dissemination and communication strategy, working hand-in-hand with the entire consortium to maximise the project's visibility and impact.
The s-X-AIPI team participated in 24 high-impact events presenting their findings and organised the “Transforming Process Industries with AI” project-dedicated concluding event on 9 April, 2025 in Belgrade. These events offered remarkable opportunities for an extensive audience reach across significant target groups worldwide including professionals from research and academia, industry, IT, software, and technology, business consulting, EU institutions, national/regional, and local authorities, as well as policy making, investors and financial stakeholders, specific end user communities, association representatives and the general public and media.
Beyond that, consortium members generated 8 open-access scientific articles (6 already published and 2 currently in the publication process), an important legacy of the project the full list of which can be found on the project website or the ZENODO repository. The AI4SAM Cluster was also formed with two more EU funded projects – AIDEAS and Circular TwAIn - to expand the project’s impact beyond individual efforts. This cross-project collaboration ranged from joint event participations in major conferences, targeted webinars and the s-X-AIPI final event to collaboration on digital communication activities to amplify each project’s outreach.
The s-X-AIPI website, designed and maintained by CORE IC, will continue as a central hub for useful information and resources. On the site, visitors can learn more about the project’s final results, important research activities performed and landmark research news of each project phase through 7 press-releases, 9 newsletter issues, videos, open-access scientific papers, public deliverables and training courses.
Additional strategic digital communication efforts included the creation of 13 videos - all available on YouTube - to convey the s-X-AIPI concept and remarkable achievements in a more engaging manner than text-based content. The project also shaped valuable online communities on LinkedInX, significantly expanding its reach, another reflection of the overall effectiveness of the D&C strategy.
For the Exploitable Results (ERs) that are closer to the market, CORE IC used its’ Profit Simulation Tool (PST) to forecast the financial revenues during the post-project commercialisation period. CORE’s PST offers a structured approach to support commercialisation planning by combining strategic business insights with market data. The financial forecasts provided by the PST were integrated effectively into the project’s exploitation strategy.
Further support from the CORE IC team
SIDENOR, a steel making facility in Spain, worked in s-X-AIPI with a focus on optimising scrap usage, especially addressing challenges from contaminants such as copper commonly found in lower-quality external scrap. Their main goal is to produce high-quality steel, prevent downstream surface defects, and minimise energy consumption in the Electric Arc Furnace (EAF) melting process. CORE IC contributed to this effort by supporting partners BFI and MSI through the development of anomaly detection software, as part of the overall solution.
Additionally, CORE IC was involved in the Asphalt use case. The aim of this use case was to target circularity of the value chain, from quarry to road, by enhancing quality control of feedstock (aggregates, bitumen, recycled asphalt), improving the overall sustainability of the production process (including asphalt paving) and the quality of final product (asphalt mix). Partners CARTIF, DEUSER, and EIFFAGE, also working on this use case, leveraged the augmented data generated by CORE IC to enhance model performance.
