Accelerator Policy: Strategy, Control, and Sustainability

• Accelerator policy is the set of formal rules, objectives, and governance mechanisms that guide the design, resource allocation, and evaluation of accelerator systems across various domains.
• It includes control strategies such as reinforcement learning for beam tuning, spill regulation, and RF setpoint optimization that use explicit state encoding, bounded actuation, and reward shaping.
• It also integrates sustainability and industrial resilience measures by establishing structured roadmaps, lifecycle assessments, and evidence-based decision frameworks for long-term operational excellence.

Searching arXiv for fresh relevant papers on accelerator policy across accelerator facilities, RL control, and accelerator management. arXiv search query: accelerator policy particle accelerator strategy reinforcement learning control accelerator management sustainability industrial base Across recent literature, the expression accelerator policy is used in several technically distinct but related senses. In large research infrastructures, it denotes the strategic governance of accelerator roadmaps, prioritization, resourcing, and risk management for facilities such as the LHC, HL-LHC, HE-LHC, future lepton colliders, and electron–ion or neutrino programs (Myers, 2011, Adolphsen et al., 2022, Barletta et al., 2014). In accelerator operations, it denotes a control law—often formulated as a policy in a Markov Decision Process (MDP)—for beamline tuning, spill regulation, or RF setpoint optimization (Ibrahim et al., 18 Oct 2025, Xu et al., 2023, Pang et al., 2020). In computing systems, it denotes a resource-allocation or scheduling rule for hardware accelerators such as GPUs, TPUs, DPUs, or heterogeneous inference engines, typically under multi-tenant, real-time, or service-level constraints (Ranganath et al., 2021, Enright et al., 2024, Zhao et al., 2024, Blanco et al., 2024, Palaniappan et al., 1 Jul 2025). A further strand treats accelerator policy as lifecycle governance for environmental sustainability and industrial capability in the design, construction, operation, and decommissioning of accelerator facilities (Wakeling et al., 24 Jan 2025, Bloise et al., 15 Sep 2025, Todd et al., 2022). Taken together, these uses define accelerator policy as the set of formal rules, objectives, constraints, and governance mechanisms by which accelerator systems are designed, allocated, controlled, and evaluated.

1. Strategic policy for accelerator facilities

In the strategic sense, accelerator policy concerns long-horizon portfolio selection across scientific frontiers, technology readiness, infrastructure reuse, and non-technical constraints. A canonical formulation appears in the CERN strategy literature, which states that the accelerator strategy “should come from the CERN Directorate,” prioritizes operation of the LHC at “7. TeV/beam up to design luminosity,” and then advances the HL-LHC “for installation in 2020-2021 and operation until around 2030” (Myers, 2011). In parallel, the strategy keeps precision-frontier options open through a Linear Collider expected “probably after the HL-LHC (>2030),” while treating “HE-LHC and neutrinos” as explicit alternatives if the Linear Collider “does not fly” for reasons including “politics, finances, governance, energy and climate situation” (Myers, 2011). That formulation makes accelerator policy an exercise in staged commitment, technology hedging, and governance under uncertainty.

The European accelerator R&D roadmap sharpens this strategic logic into topical priorities and decision gates. It organizes policy around high-field superconducting magnets, high-gradient RF systems, plasma and laser acceleration, bright muon beams and muon colliders, and energy-recovery linacs, explicitly framing the program as “timely, affordable and sustainable” (Adolphsen et al., 2022). The roadmap couples each topical area to milestones, demonstrators, resource scenarios, and future facility options, such as FCC-ee, FCC-hh, CLIC, muon-collider concepts, ERLs, and compact high-gradient systems (Adolphsen et al., 2022). Similarly, the U.S. Snowmass accelerator-capabilities synthesis frames policy around the energy frontier, intensity frontier, and QCD matter, while tying construction decisions to enabling R&D in magnets, SRF, targetry, beam diagnostics, and advanced acceleration (Barletta et al., 2014).

Several quantitative relations in these strategy documents function as policy-relevant constraints rather than merely accelerator-physics identities. For hadron rings, the bending relation $R = \frac{p}{0.3 B}$ implies that, in a fixed tunnel, higher beam momentum requires higher dipole field; this is why magnet R&D is treated as a primary policy lever for HE-LHC and future hadron colliders (Myers, 2011). For colliders more generally, luminosity depends not only on energy but also on bunch structure, beam sizes, and interaction-point optics; thus accelerator policy cannot be reduced to a single energy target (Myers, 2011, Barletta et al., 2014). For circular lepton machines, synchrotron-radiation scaling $P \propto \frac{E^4}{R}$ makes electrical power and cryogenic burden intrinsic policy variables rather than external operating details (Myers, 2011, Barletta et al., 2014). The strategic literature therefore treats accelerator policy as a constrained optimization over scientific reach, technological maturity, cost, schedule, and societal conditions.

2. Control-policy formulations in particle accelerators

In operational accelerator physics, accelerator policy is a feedback control law acting on beamline elements or machine subsystems. Recent work formulates such policies as MDPs, with explicit state, action, reward, and termination definitions. In "Reinforcement Learning for Accelerator Beamline Control: a simulation-based approach" (Ibrahim et al., 18 Oct 2025), accelerator policy is the end-to-end design of a reinforcement-learning policy that tunes beamline magnets to maximize transmission while minimizing losses on top of the Elegant simulation stack. The state is a fixed-length vector $s_t \in \mathbb{R}^{57}$ built from watch-point outputs, including statistical summaries of $(x,y,x',y')$, a normalized $5\times 5$ histogram over $(x,y)$, particle count, magnet type ID, covariance terms, and aperture parameters (Ibrahim et al., 18 Oct 2025). The action is a 4D continuous vector whose active channels depend on whether the current element is a quadrupole or dipole, with hard bounds on $K1$, $HKICK$, $VKICK$, and $FSE$ (Ibrahim et al., 18 Oct 2025).

The reward in RLABC is explicitly transmission-centric. With $P \propto \frac{E^4}{R}$0 the initial particle count, $P \propto \frac{E^4}{R}$1 and $P \propto \frac{E^4}{R}$2 counts at consecutive watch points, and $P \propto \frac{E^4}{R}$3 the number of controllable elements, the design uses

$P \propto \frac{E^4}{R}$4

If $P \propto \frac{E^4}{R}$5, the reward is $P \propto \frac{E^4}{R}$6; otherwise it is

$P \propto \frac{E^4}{R}$7

This combines global transmission, local retention shaping, and early-failure penalties (Ibrahim et al., 18 Oct 2025). The reported training setup uses DDPG with replay buffer size $P \propto \frac{E^4}{R}$8, batch size $P \propto \frac{E^4}{R}$9, discount factor $s_t \in \mathbb{R}^{57}$0, soft update $s_t \in \mathbb{R}^{57}$1, actor learning rate $s_t \in \mathbb{R}^{57}$2, critic learning rate $s_t \in \mathbb{R}^{57}$3, and Gaussian exploration noise (Ibrahim et al., 18 Oct 2025). On two beamlines, the policy reaches transmission rates of $s_t \in \mathbb{R}^{57}$4 and $s_t \in \mathbb{R}^{57}$5, comparable to expert tuning (Ibrahim et al., 18 Oct 2025).

A second formulation appears in Mu2e spill regulation. In "Beyond PID Controllers: PPO with Neuralized PID Policy for Proton Beam Intensity Control in Mu2e" (Xu et al., 2023), accelerator policy is a learned feedback law that regulates the Spill Regulation System on millisecond timescales. The control objective is to drive the spill intensity $s_t \in \mathbb{R}^{57}$6 toward a reference $s_t \in \mathbb{R}^{57}$7 over a $s_t \in \mathbb{R}^{57}$8 ms spill sampled at $s_t \in \mathbb{R}^{57}$9 kHz, giving $(x,y,x',y')$0 control steps per episode (Xu et al., 2023). The state consists of proportional, integral, and derivative error terms plus the previous action:

$(x,y,x',y')$1

with $(x,y,x',y')$2 ms (Xu et al., 2023). The reward is the negative exponential moving average of the absolute tracking error, with $(x,y,x',y')$3 in the reported experiments (Xu et al., 2023). The neuralized PID actor implements

$(x,y,x',y')$4

and is trained with PPO. The paper reports an average Spill Duty Factor improvement of $(x,y,x',y')$5 over unregulated spills and an additional $(x,y,x',y')$6 over the current PID baseline across nine random spill scenarios (Xu et al., 2023).

An earlier deep-RL accelerator-control example targets RF cavity tuning in a drift tube linac (Pang et al., 2020). There the policy is a Gaussian continuous-control policy trained with A3C to adjust five RF setpoints, using a 15-dimensional state built from currents, loss-power monitors, and absolute control-variable values (Pang et al., 2020). The reward assigns $(x,y,x',y')$7 when the downstream current fraction satisfies $(x,y,x',y')$8, and otherwise penalizes deviations from target transmission and beam-loss power with downstream-increasing weights (Pang et al., 2020). The 3-action configuration consistently reaches the target from random initial points within 300 steps, while the 5-action configuration reaches an 83% success rate from a fixed start after multi-agent asynchronous exploration (Pang et al., 2020).

These works show that, in accelerator control, policy is not a metaphor for management. It is a mathematically specified controller with state encoding, bounded actuation, reward shaping, and termination logic. A common misconception is that such policies are synonymous with unconstrained optimization. The reported systems instead rely on clipped actuation, channel masking, per-step limits, failure termination, and, for real deployment, proposed safety layers, rate limiters, shadow-mode validation, rollback, and human oversight (Ibrahim et al., 18 Oct 2025, Xu et al., 2023, Pang et al., 2020).

3. Accelerator policy in computational resource management

A distinct literature uses accelerator policy to denote resource arbitration over GPUs, TPUs, and heterogeneous accelerators in shared systems. The technical issue is no longer beam transport but contention, topology, latency, and quality of service.

In "MAPA: Multi-Accelerator Pattern Allocation Policy for Multi-Tenant GPU Servers" (Ranganath et al., 2021), the server topology is modeled as a weighted graph $(x,y,x',y')$9 whose edges carry bandwidth and latency attributes, while each workload is another graph $5\times 5$0 whose edge weights encode communication intensity. The placement problem seeks an injective mapping $5\times 5$1 maximizing a bandwidth/latency-aware objective. MAPA enumerates subgraph embeddings with Peregrine and scores them using Aggregated Bandwidth, a learned Effective Bandwidth model, and Preserved Bandwidth for future jobs (Ranganath et al., 2021). The policy is explicitly preservation-aware: bandwidth-sensitive jobs receive the placement with highest predicted Effective Bandwidth, while bandwidth-insensitive jobs receive the placement with highest Preserved Bandwidth (Ranganath et al., 2021). On DGX-1 V100, the MAPA Preserve policy reduces 75th-percentile execution time by $5\times 5$2, reduces worst-case execution time by up to $5\times 5$3, and improves throughput by $5\times 5$4 relative to an ID-based placement baseline (Ranganath et al., 2021).

PAAM addresses a different computational setting: coordinated, priority-driven accelerator access in multi-process ROS 2 robotics (Enright et al., 2024). Here a standalone ROS 2 executor acts as an accelerator resource server, maintains one accelerator context, and arbitrates requests by chain priority using a two-level hierarchy of priority buckets and per-bucket queues (Enright et al., 2024). The framework provides worst-case response-time analysis for callback chains whose accelerator segments execute non-preemptively within a bucket and may be preempted across buckets if the device supports it (Enright et al., 2024). The paper reports that, in accelerator-heavy robotic systems, PAAM may achieve up to a 91% reduction in end-to-end response time of critical callback chains, with a highlighted reduction from 917 ms to 83 ms in realistic autonomous-driving scenarios (Enright et al., 2024).

Arcus casts cloud accelerator policy as traffic management rather than mere compute allocation (Zhao et al., 2024). Its control plane maintains per-flow status, profiled capacity models, and token-bucket parameters, while an accelerator-side interposition layer reshapes traffic across all paths. This policy targets predictable accelerator performance under communication-induced contention in PCIe complexes, DMA engines, buffers, and NIC paths (Zhao et al., 2024). Reported outcomes include tail-latency reductions from 128/193/299 $5\times 5$5s at the 95th/99th/99.9th percentiles to 104/133/162 $5\times 5$6s, i.e. $5\times 5$7, $5\times 5$8, and $5\times 5$9, with throughput variance below 1% (Zhao et al., 2024).

RELMAS addresses online DNN scheduling on heterogeneous multi-accelerator systems using DDPG (Blanco et al., 2024). Its state encodes queue contents, waiting times, deadlines, and per-accelerator compute and bandwidth descriptors for each layer, while the action specifies a temporal priority and an accelerator score vector per sub-job (Blanco et al., 2024). The reward combines SLA satisfaction with normalized slack, and the simulator explicitly models shared-DRAM bandwidth contention (Blanco et al., 2024). The reported result is up to a 173% improvement in SLA satisfaction rate over prior schedulers, with less than 1.5% energy overhead (Blanco et al., 2024).

HERCULES moves accelerator policy itself into hardware. It implements stochastic online scheduling in an FPGA using a modified greedy cost-selection policy over heterogeneous machines, combining WSPT-ordered virtual schedules and a release point $(x,y)$0 (Palaniappan et al., 1 Jul 2025). With INT8 quantization and parallel cost computation, it achieves up to 1060x speedup over a single-threaded software scheduler while using up to about 21 W and up to 13% of FPGA resources (Palaniappan et al., 1 Jul 2025). In this context, accelerator policy is a hardware-realized scheduling mechanism.

A plausible implication is that the computational literature converges on a common structure despite highly different environments: explicit system models, contention-aware state, bounded or prioritized actions, and objective functions that trade off latency, throughput, fairness, or SLA satisfaction under heterogeneity.

4. Governance, safety, and deployment constraints

Accelerator policy, whether for beam control or accelerator allocation, is constrained by governance. In facility strategy, governance specifies who sets priorities, how readiness is judged, and which milestones trigger commitments. CERN strategy assigns primacy to the CERN Directorate and uses formal studies, workshops, and external community support as inputs (Myers, 2011). The European R&D roadmap organizes work under Lab Directors Group oversight with milestones in 2025, 2027, and 2030–2031 to support evidence-based technology selection (Adolphsen et al., 2022). Snowmass-style U.S. planning similarly ties project sequencing to explicit technological capability and program balance across frontiers (Barletta et al., 2014).

In control-policy deployment, governance appears as safety validation, operator oversight, and rollback logic. The RLABC study notes that its simulation does not include hardware ramp rates or interlocks, and states that real deployment should add safety layers and rate limiters (Ibrahim et al., 18 Oct 2025). It further recommends shadow-mode testing, certified bounds, action filters, rollback to expert or last-known-safe configurations, and operator approval for large deviations (Ibrahim et al., 18 Oct 2025). Mu2e control experiments maintain hardware limits and fallback to a conservative PID when anomalies are detected (Xu et al., 2023). The DTL control work enforces action ranges and per-step increments to avoid protection-system trips and adds an action-bound regularizer to the policy loss (Pang et al., 2020).

Computational accelerator governance similarly centers on predictable arbitration. PAAM uses admission control based on worst-case response-time analysis, maps chain criticality into accelerator-priority buckets, and distinguishes between suspension and spinning because these choices alter interference terms in the schedulability model (Enright et al., 2024). Arcus uses admission control, path selection, and proactive reshaping to keep flows within profiled SLO-friendly operating regions (Zhao et al., 2024). HERCULES exposes tunables such as $(x,y)$1, schedule capacity $(x,y)$2, and admission overlays that can be layered on top of the base greedy policy (Palaniappan et al., 1 Jul 2025).

One recurrent controversy is whether learned policies are inherently opaque and therefore unsuitable for accelerators. The cited work does not treat opacity as a reason to reject policy-based control; instead it treats interpretability, auditable logging, certified bounds, and fallback mechanisms as deployment requirements (Ibrahim et al., 18 Oct 2025, Xu et al., 2023, Enright et al., 2024). Another misconception is that accelerator governance applies only at the facility level. The literature shows governance operating at all scales: portfolio selection, subsystem certification, control-stack interlocks, and per-request arbitration.

5. Sustainability and industrial policy

Recent accelerator policy literature extends beyond performance and governance to environmental sustainability and industrial capability. "High-level environmental sustainability guidelines for large accelerator facilities" (Wakeling et al., 24 Jan 2025) frames sustainability as a lifecycle issue spanning planning, construction, operation and maintenance, and decommissioning. It calls for alignment with the Paris Agreement and related commitments, use of life-cycle assessment under ISO 14040/14044, and transparent accounting across Scope 1, 2, and 3 emissions (Wakeling et al., 24 Jan 2025). The guidance emphasizes prevention and reduction before offsets, hotspot analysis during optioneering, responsible procurement, sub-metering, demand shifting, waste-heat recovery, helium recovery, low-carbon materials, and climate adaptation (Wakeling et al., 24 Jan 2025).

The LDG Sustainability Working Group broadens this into a formal assessment framework for future accelerators (Bloise et al., 15 Sep 2025). It defines the functional unit as the full research infrastructure, including accelerators, detectors, and technical support systems, and adopts lifecycle stages A1–C4 together with greenhouse-gas accounting and cost-benefit analysis including environmental externalities (Bloise et al., 15 Sep 2025). It recommends common LCIA methods such as ReCiPe 2016 midpoint (H), EPD requirements, renewable-energy procurement, heat-recovery plans, and reporting under GRI, ESRS, or EMAS (Bloise et al., 15 Sep 2025). It also gives policy-level formulas for CO$(x,y)$3e conversion, discounted net shadow cost of carbon, and net present value (Bloise et al., 15 Sep 2025).

Industrial policy is treated separately in the U.S. white paper on the accelerator technology base (Todd et al., 2022). That work argues that the United States lacks a sufficiently robust domestic vendor base for critical accelerator technologies and documents barriers arising from procurement practice, regulatory burden, limited early engagement with industry, and a technology-transfer model that prioritizes licensing over structured knowledge transfer (Todd et al., 2022). Case studies in superconducting RF cavities, undulators, and accelerator-grade magnets are used to show how domestic capability weakened while European and Japanese ecosystems were strengthened by infrastructure subsidies, sustained lab–industry co-development, and preferential local procurement (Todd et al., 2022). The recommendations include risk-appropriate contracts, pre-RFP design workshops, co-funded industrial infrastructure, loaned metrology equipment, standards programs, and SBIR/STTR alignment with future procurements (Todd et al., 2022).

These sustainability and industrial strands modify the meaning of accelerator policy. The subject is no longer only how to build or operate accelerators, but how to make them socially licensable, resilient, and industrially executable over multi-decade lifecycles. This suggests that accelerator policy now functions as an integrative framework in which energy efficiency, embodied carbon, workforce, procurement, and scientific capability are co-equal decision variables.

6. Conceptual unification and future directions

The literature does not present a single universal theory of accelerator policy, but it does reveal a shared architecture. First, policy is always tied to an explicit objective: energy-frontier leadership, beam transmission, spill uniformity, SLA satisfaction, bounded worst-case response time, reduced tail latency, lower embodied carbon, or domestic industrial resilience (Myers, 2011, Ibrahim et al., 18 Oct 2025, Xu et al., 2023, Enright et al., 2024, Zhao et al., 2024, Bloise et al., 15 Sep 2025, Todd et al., 2022). Second, policy is always constrained: by magnet fields, aperture, interlocks, reward shaping, device topology, process priorities, traffic paths, environmental limits, or procurement law. Third, policy is always implemented through a formal mechanism: strategic roadmaps, MDPs, graph mappings, token buckets, priority queues, WCRT equations, LCA/CBA frameworks, or industrial-development programs.

Several future directions recur across the sources. In beam control, broader RL algorithms such as PPO and SAC, constrained RL, offline RL, model-based RL, and multi-objective objectives including emittance preservation and Twiss targets are proposed as next steps (Ibrahim et al., 18 Oct 2025). In Mu2e control, future work includes constrained RL, explicit stability penalties, and more realistic multi-input control of the three quadrupole currents (Xu et al., 2023). In computational systems, learned or hardware-accelerated schedulers are increasingly combined with explicit contention modeling, safety analysis, or SLO-aware telemetry (Blanco et al., 2024, Palaniappan et al., 1 Jul 2025). In facility policy, R&D roadmaps, sustainability frameworks, and industrial-base strategies are converging on common themes of evidence-based gates, standardized metrics, and lifecycle accountability (Adolphsen et al., 2022, Bloise et al., 15 Sep 2025, Todd et al., 2022).

A persistent source of confusion is the term itself. In one body of work it denotes institutional strategy; in another, a control policy $(x,y)$4; in another, a scheduler or access manager. The cited literature shows that these are not unrelated homonyms. Each concerns the formal selection of actions under constraints in accelerator-centered systems. The scale differs—from cavities and magnets to laboratories and cloud clusters—but the policy function remains the same: to translate objectives, models, and constraints into operational decisions that are technically effective and institutionally governable.