Research Authors & Affiliations

Mithila Arman¹, Md. Farhad Kabir², Fakir Mashuque Alamgir³, Kazi Aklima⁴, Mohammad Kasedullah⁵, Efaz Kabir⁶, Abdullah Rakib Akand⁷, Abdullah Al Imtihan⁸, & Miguel Ángel Rodríguez Sánchez⁹

¹Dept. of Computer Science and Engineering, BRAC University, Dhaka, Bangladesh
²Marshall School of Business, University of Southern California, Los Angeles, CA, USA
³Dept. of Electrical and Electronic Engineering, East West University, Dhaka, Bangladesh
⁴Dept. of Mathematics, Jagannath University, Dhaka, Bangladesh
⁵Dept. of Computer Science and Engineering, Varendra University, Rajshahi, Bangladesh
⁶Dept. of Computer Science and Engineering, East West University, Dhaka, Bangladesh
⁷Dept. of Computer Science and Engineering, Asian University of Bangladesh, Dhaka, Bangladesh
⁸Dept. of Artificial Intelligence, European Institute for Materials AI & Technology, Madrid, Spain
⁹Dept. of Computer Science and Engineering, University of Salamanca, Madrid, Spain

Journal Information

Journal: Machine Learning (Springer Nature)
Timeline: Received: 06 Feb 2026 | Revised: 18 May 2026 | Accepted: 02 June 2026
Published / Record Date: 15 June 2026
DOI: 10.1007/s10994-026-07095-x

The exponential proliferation of scientific literature presents formidable impediments to comprehensive understanding and strategic anticipation of future research trajectories across diverse domains. While extant computational methodologies, including bibliometrics and static topic modeling, facilitate the analysis of prevailing trends and thematic structures, the rigorous forecasting of the complex structural evolution inherent in the diachronic development of scientific knowledge remains a largely unresolved challenge. Addressing this lacuna, we introduce DynSciGraph, a computational framework engineered to harness the semantic interpretation capabilities of Large Language Models (LLMs) for the high-fidelity extraction of fine-grained conceptual entities and semantic relations from extensive corpora of scholarly publications. The framework subsequently leverages Temporal Graph Neural Networks (TGNNs) to model the non-linear dynamics intrinsic to time-evolving, domain-specific knowledge graphs (KGs), coupled with sequence models for multi-horizon predictive tasks. A distinctive architectural feature of DynSciGraph is its capacity to integrate heterogeneous external signals encompassing bibliometric indicators, research funding allocation patterns, and patenting activities, thereby augmenting the contextual richness of its forecasts. The evidence reported in this manuscript is exclusively simulation-based and does not yet demonstrate real-world forecasting performance. We present a detailed exposition of the architectural design and report proof-of-concept simulation results on synthetic corpora parameterized to reflect characteristics of Materials Science and Artificial Intelligence domains. All reported quantitative outcomes derive from a controlled simulation environment whose generative process is fully specified in Sect. 4.1; no measurements on real longitudinal scientific corpora are reported, and comprehensive empirical validation on real data constitutes ongoing and future work. Within this simulation, the full DynSciGraph configuration achieves higher mean AUC for link prediction (up to 0.86), higher Hit Rate@50 for emergent-entity identification, and higher Spearman rank correlation for hotspot prediction than the implemented baselines; because the simulation’s generative process embeds the same inductive biases the architecture is designed to exploit, these orderings reflect properties of the generative process and should not be interpreted as evidence of real-world superiority. This investigation delineates a computationally grounded paradigm for scientific foresight that, pending empirical validation, holds promise for proactive identification of latent research lacunae, discovery of synergistic interdisciplinary connections, and more strategic planning of scientific inquiry.