Desiree Ruiz

With the goal of facilitating efficient access to data to support the development of new R&D&I initiatives, this paper presents an information retrieval RAG-based system for projects funded by the Horizon 2020 program domain through multiple configurations and specific adaptations, considering CPU-only hardware demand. This work is an extension of a previous proposal in order to address the limitations when moving from exploratory prototypes to robust, large-scale experimentation. Additionally, we propose an automatic evaluation pipeline that enables large-scale, reproducible comparison of different configurations by analyzing generated answers and automating the scoring process. A methodology structured into three main phases guided the development of the system. First, project embeddings were generated using vector representation techniques, segmenting documents to compare models in terms of semantic quality. Second, relevant fragment retrieval was performed using language models to assess their relevance to a query. Third, system performance was evaluated using language model-based evaluators, automating the process. Additionally, an in-depth analysis identified evaluators that best approximate human judgment, comparing their results with metrics from the reference study. This work contributes to the State of the Art by proposing improvements in information retrieval systems.

Recent studies have highlighted the potential of Large Language Models (LLMs) to become experts in specific areas of knowledge through the utilization of techniques that enhance their context. Nevertheless, an interesting and underexplored application in the literature is the creation of an LLM that specializes in research projects, as it could streamline the process of project ideation and accelerate the advancement of research initiatives. In this regard, the aim of this work is to develop a tool based on LLM technology capable of assisting the employees of technology centers in answering their queries related to research projects funded under the Horizon 2020 program. By facilitating the identification of suitable funding calls and the formation of consortia with partners meeting specific requirements, tasks that are traditionally time-intensive, the proposed tool has the potential to improve operational efficiency and enable technology centers to allocate their resources more effectively. To improve the model’s baseline performance, context extension techniques such as Retrieved Augmented Generation (RAG) and prompt engineering were explored. Specifically, different RAG approaches and configurations, along with a specialized prompt, were tested on the LLaMA 3 70B model, and their results were compared to those obtained without context extension. The proposed evaluation metrics, which aligned with human judgment while maintaining objectivity, revealed that RAG systems outperformed the standalone LLaMA 3 70B, achieving a rate of optimal responses of up to 46% compared to 0% for the baseline model. These findings emphasize that integrating RAG and prompt engineering pipelines into LLMs can address key limitations, such as generating accurate and informative answers. Moreover, this study demonstrates the practical feasibility of leveraging advanced LLM configurations to support research-driven organizations, highlighting a pathway for the further development of intelligent tools that enhance productivity and foster innovation in the research domain.

This paper presents the results of a study on data preprocessing and modeling for predicting corrosion in water pipelines of a steel industrial plant. The use case is a cooling circuit consisting of both direct and indirect cooling. In the direct cooling circuit, water comes into direct contact with the product, whereas in the indirect one, it does not. In this study, advanced machine learning techniques, such as extreme gradient boosting and deep neural networks, have been employed for two distinct applications. Firstly, a virtual sensor was created to estimate the corrosion rate based on influencing process variables, such as pH and temperature. Secondly, a predictive tool was designed to foresee the future evolution of the corrosion rate, considering past values of both influencing variables and the corrosion rate. The results show that the most suitable algorithm for the virtual sensor approach is the dense neural network, with MAPE values of (25 ± 4)% and (11 ± 4)% for the direct and indirect circuits, respectively. In contrast, different results are obtained for the two circuits when following the predictive tool approach. For the primary circuit, the convolutional neural network yields the best results, with MAPE = 4% on the testing set, whereas for the secondary circuit, the LSTM recurrent network shows the highest prediction accuracy, with MAPE = 9%. In general, models employing temporal windows have emerged as more suitable for corrosion prediction, with model performance significantly improving with a larger dataset.

The work describes the use of Electrochemical Impedance Spectroscopy (EIS) and Equivalent Circuit Models (ECMs) for predicting the aging of lithium-ion (Li-ion) batteries, in order to accelerate the assessment of battery cells used in space applications. The aging of Li-ion cells can lead to capacity fade, power fade, and safety issues, making it essential to develop techniques for predicting their performance under different operating conditions. The use of machine learning (ML) techniques is also considered crucial for battery research and development due to the high number of variables involved in the problem and the massive amount of data generated daily. The methodology involved in analysing EIS data, selecting suitable datasets and exploiting them to develop representative mathematical models for the cell status at each studied age stage is described here. The text also discusses the use of statistical and machine learning methods for understanding and predicting the state of health of Li-ion batteries based on specific parameters. The results of the study showed that the Cambridge dataset was the most suitable for the analysis of the aging process. On the other hand, the Support Vector Machine for Regression (SVR) and the K-Nearest Neighbors Regression (KNN-Re) methods were selected for training a contrasted correlation model on the study dataset. The feature importance calculation using the permutation importance method revealed that the test temperature was the most important feature for predicting the state of health, followed by Warburg impedance.