Today I submitted PSI-MEM (Physics Structure-Informed Modeling of Printed Memristors for Neuromorphic and Intelligent Sensing), my own very first research proposal built around a research direction that I developed and now want to take into an experimental setting.
The proposal was submitted to the Marie Skłodowska-Curie Actions Postdoctoral Fellowships (MSCA-PF) call, with CENIMAT|i3N at NOVA FCT as the host institution and Dr. Emanuel Carlos as supervisor. The 2026 call closes on 9 September 2026, and this year also marks the 30th anniversary of the Marie Skłodowska-Curie Actions, which began in 1996. Submitting my first MSCA-PF proposal in this anniversary year gives the occasion an additional personal significance.
This is where several apparently separate parts of my research trajectory finally converge. My doctoral work focused on sustainable deep learning systems. At the German Aerospace Center, I worked on physics-guided learning for autonomous systems. At TU Chemnitz, I encountered printed memristors and three-dimensional neuromorphic circuits. A particularly important turning point came in 2025, when I discovered the Ψ-NN work of Liu et al. The idea that a physics-informed model could be used not only to learn parameters but also to discover structure became the conceptual starting point for the Ψ-family methods (Ψ-HDL, Ψ-xLSTM, and Ψ-Vortex) I subsequently developed during an independent research period after returning back from ERCEA. Together, these methods developed different parts of a physics structure-informed workflow spanning structure inference, recurrent device modeling, structure selection, and Verilog-A compact-model generation.
Existing memristor modeling approaches span physics-based compact models, phenomenological descriptions, and increasingly data-driven surrogates. Physics-based models can provide interpretable device descriptions, but they often require assumptions about switching mechanisms and internal states that are difficult to validate in printed devices. Flexible data-driven models can reproduce measured behaviour, but predictive accuracy alone does not establish whether the learned representation captures reproducible device dynamics or whether it can be translated into a compact form suitable for circuit simulation. PSI-MEM is positioned between these two extremes: it asks whether structural regularities can be inferred from experimental device populations, tested for reproducibility under fabrication-induced variability, and converted into compact executable digital twins.
PSI-MEM asks what happens when those methods leave carefully controlled computational studies and meet real experimental devices. The central question is simple to state but difficult to answer: Do printed memristors contain structural regularities that remain reproducible despite cycle-to-cycle, device-to-device, and fabrication-batch variability? And, if they do, can that structure be turned into compact digital twins that work inside circuit simulators?
PSI-MEM also sits within a broader European research effort around memristive and neuromorphic technologies. Recent MSCA-PFs have supported projects including MemSusPer, focused on sustainable memristors for neuromorphic computing, MALEFICENT, combining memristive devices with meta-learning, EASIFeT, connecting electron-device modeling and simulation with artificial synapses, and EEHIMIC, addressing memristor-based in-memory computing and hardware implementation. More recently, FLECS-2D was funded to investigate memristive devices for in-sensor neuromorphic computing. PSI-MEM approaches this research landscape from a different direction. Its central objective is not primarily to develop a new memristor technology or neuromorphic architecture, but to determine whether experimentally observed printed-memristor dynamics contain reproducible structural organization that can be translated into variability-aware compact digital twins for circuit simulation.
One aspect of the proposal that became increasingly important to me while writing it is that the answer does not have to be yes. A sufficiently flexible ML model can always find patterns. That does not mean those patterns correspond to something reproducible in the device population. PSI-MEM is therefore designed around a falsifiable decision. A common structure may exist, structure may only be stable within particular devices or operating regimes, or no single stable structure may be supported at all. In the latter case, variability itself becomes the scientific result.
That way of thinking changed how I approached the proposal.
That was probably the most valuable part of preparing PSI-MEM. The collaboration with CENIMAT|i3N is central to that logic. I bring the computational side, including the Ψ-family, structure inference, Verilog-A generation, and SPICE validation. The host brings the experimental capability I do not yet have at the same level, namely: inkjet-printed memristor fabrication, electrical characterization, and materials-level interpretation of switching behaviour. If funded, the goal is to connect those capabilities from end to end: fabrication → measurement → structure inference → variability testing → digital twins → Verilog-A → circuit-level co-simulation.
There is also something personally significant about submitting this proposal. Until now, much of my recent work has been about establishing individual pieces of the methodology. Ψ-HDL addressed structure-to-Verilog-A generation, and Ψ-xLSTM extended the approach to recurrent dynamics. Ψ-Vortex moves further toward structure-regularized recurrent inference. PSI-MEM is the first time I have tried to turn that sequence into a complete research programme with its own experimental question, training plan, risks, milestones, and longer-term scientific direction.
Later Edit: The European Commission has published on 11 September 2026 the submission figures for the 2026 MSCA Postdoctoral Fellowships call. A record 21,627 proposals were submitted, 26.7% more than in 2025 and the highest number ever received by an individual funding action across the 40-year history of the EU’s research and innovation framework programmes. Of these, 20,131 are European Fellowship proposals and 1,496 are Global Fellowship proposals. The call currently has an indicative budget of €399.05 million, expected to fund approximately 1,600 projects. Host organisations in Portugal, where PSI-MEM was submitted, coordinate 2.99% of all proposals.
These figures make clear just how competitive this year’s evaluation will be. At the same time, they place PSI-MEM within an extraordinary moment for European research: a record level of participation coinciding with the 30th anniversary of the MSCA. The evaluation will run until mid-February 2027, when applicants are expected to receive their results. Until then, I remain glad to have contributed my own research direction to this remarkable call.
Neueste Kommentare