EnableFusion | May 29, 2026
The question facing fusion is changing.
For decades, the central question of fusion research has been "Can we create a plasma?" And now, around the world, the answer is converging on "yes." ITER's assembly is underway, private fusion companies are operating demonstration devices, and AI-based real-time plasma control is being validated not just in the lab but on actual machines.
But the next question that fusion must answer as a commercial energy source lies on an entirely different level: "Can we operate the plasma 24 hours a day, 365 days a year, economically?" This question cannot be answered by hardware alone.
The SDF (Software-Defined Fusion) White Paper we are releasing today is EnableFusion's answer to that question.
SDF is a concept that applies the SDV (Software-Defined Vehicle) paradigm established in the automotive industry to the fusion domain. Its goal is to transform tokamak operation — which has evolved with a hardware-centric focus — into a paradigm in which AI software takes the lead in control.
Today, the core competency of fusion operation exists as tacit knowledge held in the experience and intuition of a small number of experts. This knowledge is difficult to document, difficult to reproduce, and difficult to transfer to other machines. SDF systematizes this tacit knowledge within AI models and digital twins, transforming it into an organizational asset that is not dependent on any single individual. On top of that, when an operator simply inputs a goal, the AI designs the scenario and performs control on a millisecond timescale — this so-called 'Full Self-Driving for Fusion' is SDF's ultimate destination.
Among the design principles of SDF, I want to emphasize one of the most important. SDF does not replace existing fusion control infrastructure such as ITER CODAC or EPICS. It is a non-invasive extension that layers an AI-Native operation tier as a 'plug-in overlay' on top of hardware and middleware proven over decades.
This approach is intentional. Adding new value on top of the engineering assets the fusion community has accumulated, while respecting them, is what we believe to be realistically the fastest and safest path.
The white paper begins by defining the concept of SDF and then systematically develops the following.
It defines the architecture of the three core layers that make up SDF — a Fusion LLM-based user interface, an FSD Multi-Agent system, and an Engine Digital Twin responsible for real-time predictive simulation — and explains how they cooperate to realize autonomous operation through three concrete operating scenarios.
Regarding reliability and safety, it presents a four-layer safety architecture: the way physical constraints are embedded into the AI structure (Physics-Informed ML), anomaly detection and automatic fallback, AI-based virtual diagnostics, and explainability. "Using AI" is not enough. Without a design for "what happens when the AI fails," AI cannot earn trust in commercial fusion operation.
Furthermore, it proposes Backward Design as the long-term vision of SDF. Until now, fusion device design has followed a "build the maximum possible and experiment" approach. Once SDF is sufficiently mature, this direction reverses. It becomes possible to "first define the target performance, then back-calculate the minimum hardware specifications needed to achieve it." We believe this is the most essential role software can play in the paradigm shift from R&D to a commercial energy source.
The field of fusion AI control is moving fast. DeepMind and EPFL's real-time reinforcement learning control on TCV (2022), Princeton's instability avoidance on DIII-D (2024), the TORAX simulator from CFS and Google, the digital twin from NVIDIA and General Atomics — progress at each individual layer is already substantial.
However, there is still no software platform that integrates these pieces under the single goal of commercial power plant operation. We judged that this integration is something someone must inevitably do, and that now is the time to share that vision. By first releasing our thinking on direction and architecture — rather than a finished product — we want to begin a dialogue with the fusion community.