普林斯顿大学研究人员开发了名为PACMAN的人工智能系统，能够在毫秒级时间内监测和控制聚变等离子体，其反应速度远远超过人类操作员的能力1。该系统已在美国圣地亚哥的DIII-D国家聚变设施进行了五次实验验证，展现了在核聚变领域应用人工智能的重大进展。
PACMAN框架在约20毫秒内完成一个完整的控制循环1，相比人类操作员通常需要数秒的反应时间快得多。在其中一次实验中，AI模型提前约200毫秒预测到撕裂模式不稳定性，并成功阻止其形成1。系统能够实时协调DIII-D设施的全部6个回旋管，通过调整功率和镜位参数达到控制目标1。
研究团队保留了人类在系统目标制定中的决策权，同时确保严格的安全控制得以维持1。值得注意的是，第二个AI模型的部署仅需数天时间即可完成，相比首个模型耗时数个月的开发周期大幅缩短1。该研究成果已发表在《核聚变》期刊2026年第66卷第7期1。
Researchers at Princeton University have developed an artificial intelligence system called PACMAN that monitors and controls fusion plasma on millisecond timescales, far outpacing the reaction speed of human operators.1 The system completed one full control cycle in approximately 20 milliseconds, dramatically faster than the second-level responses typical of human control.1
The PACMAN framework was successfully deployed across five experimental runs at the DIII-D National Fusion Facility in San Diego.1 In one instance, the AI model predicted a dangerous tearing mode instability approximately 200 milliseconds in advance and successfully prevented its formation.1 The system also demonstrated the ability to coordinate all six of the facility's gyrotrons in real time, dynamically adjusting power and mirror positions to achieve target parameters.1 Importantly, the design maintained strict safety controls and preserved human decision-making authority over the system's objectives.1
Development efficiency also improved significantly with experience. While the initial model required several months to implement, installation of a second model was completed in just days.1 The research has been published in the journal Nuclear Fusion, Volume 66, Issue 7, 2026.1
评论
还没有评论，欢迎留下第一条。