<p>This research work presents a novel IR system, which employs a multi-scale-based deep learning model with a heuristic algorithm to address the issues in this experimental field. Initially, the required data is collected manually, which comprises scientific Portable Document Format (PDF) papers with different topics. Further, the PDF file is converted into docx format, whereas multiple details like author, subsections, title, abstract, table, figure, etc., are extracted. During training process, the essential features are identified from the extracted data by parameter optimized multi-scale adaptive transformer network (PO-MATN), in which some of the parameters are optimized using an improved chameleon swarm algorithm (ICSA). These meaningful features enhance the model’s efficacy and also performance rate. These extracted features are further transformed into a more informative form using weighted feature pool formation and are stored in the feature library. During testing stage, once the user sends the query, the MATN is utilized to determine the query features. Finally, the similarity check is performed for both the trained and test features to retrieve the related documents. Thus, the designed framework effectively performs the IR process from scientific publications. Finally, the experiments of this designed framework are conducted by analyzing traditional approaches. From the experimental findings, the developed model shows 85%, 86% and 86.5% precision, recall and F1-score values, thus ensuring the efficacy of the suggested IR process.</p>