宽视场射电暴望远镜设计工具Wide-field radio burst telescope design tool L 波段 · 快速射电暴与银河系磁星 · 专家主导、AI 辅助的设计案例L band · fast radio bursts and Galactic magnetars · an expert-led, AI-assisted design study

输入预算与科学目标，得到一台宽视场射电暴望远镜的完整构型Enter a budget and science goals, get a complete design for a wide-field radio burst telescope

口径、单元数、采集板与 FPGA 板卡数量、波束数、灵敏度、亮暴年产率和成本分项，随输入实时更新。Aperture, element count, acquisition and FPGA boards, beams, sensitivity, bright-burst yield and cost breakdown update as you change the inputs.

简洁版只留 5 个关键输入；详细版开放全部价格与技术参数，并给出数字链路、预算曲线与三档对照。所有结果都是模型预测The simple view keeps five key inputs; the detailed view opens every price and technical parameter and adds the digital chain, budget curves and a three-tier comparison. All results are model predictions PRED，不是测量。, not measurements.

系统示意：竹节阵面 → 阵面下方竖插的 XCZU67DR 采集板 → 光纤 → 100G 交换机 → KU5P 集群与服务器；选择外站构型时，下方同比例画出外站及其与主站的广域网连接。数量随当前设计实时更新；板卡与机柜不按比例。System sketch: bamboo aperture → XCZU67DR acquisition boards mounted vertically beneath it → optical fibre → 100G switches → KU5P cluster and servers; with outriggers, they are drawn below at the same scale with their wide-area link to the main station. Counts follow the current design; boards and racks are not to scale.

从科学目标到硬件From science goals to hardware PRED

阵面排布（示意）Aperture layout (sketch) DESIGN

成本分项Cost breakdown PRED

亮 FRB 年产率随预算变化Bright FRBs per year versus budget PRED

实线：亮端斜率 −1.5；虚线：−2.2；阴影：系统温度 50–100 K 与两种斜率的情景极值。圆点为当前设计。Solid: bright-end slope −1.5; dashed: −2.2; shading: scenario extremes for system temperatures of 50–100 K and both slopes. Dots mark the current design.

磁星目标Magnetar objective PRED

按科学侧重推荐指向与视场Pointing and field advised by the science emphasis PRED

数字链路Digital chain DESIGN PRED

三档预算对照（当前参数与构型）Three budget tiers (current parameters and configuration) PRED

完整输出Full output

我们的设计：专家主导、AI 辅助Our design: expert-led, AI-assisted

这个工具来自一项完整的设计工作。团队具有地面射电干涉阵的建造与运行经验，确定科学目标、规格与架构，确认器件价格并批准每一项修改；两个 AI agent 在团队指导下承担执行与建模：设计 agent（OpenAI Codex）操作全波电磁、电路求解器与 FPGA 工具链，规划 agent（Anthropic Claude）建立科学产出与成本模型。两个模型家族互相审阅，全部会话留有日志。This tool comes from a complete design study. The team, experienced in building and operating ground-based radio interferometers, set the science goals, specifications and architecture, confirmed component prices and approved every change. Two AI agents did the execution and modelling under the team's direction: a design agent (OpenAI Codex) operated full-wave electromagnetic and circuit solvers and the FPGA toolchain, and a planning agent (Anthropic Claude) built the science-yield and cost models. The two model families reviewed each other's work, and every session was logged.

B2 bamboo element render

天线单元：B2 圆肩竹节Element: B2 “bamboo” outline DESIGN

车削回转体，高 158 mm、直径 108 mm，阵距 112.5 mm；相邻竹节体之间的渐变缝隙构成双极化辐射缝，底部净距 4.5 mm。A lathe-turned body of revolution, 158 mm tall and 108 mm across, on a 112.5-mm pitch; the tapered gaps between neighbouring bodies form the dual-polarized radiating slots, 4.5 mm wide at the base.

电磁性能待全波仿真验证Electromagnetic performance awaits full-wave simulation PENDING

Aperture render

阵面：第0代网格上的 1,000 个单元Aperture: 1,000 elements on the Gen-0 grid DESIGN SIM

25 个 40 单元模块、1,928 路输入；图为待建造的竹节单元。全波仿真数据来自原单元外形：800 个固定波束铺满 ±40° 视场，0.9–1.4 GHz 最坏扇贝损耗 1.96 dB（锥削代价 0.7 dB）。25 modules of 40 elements and 1,928 inputs, rendered with the bamboo element to be built. Full-wave results are for the previous element outline: 800 fixed beams tile the ±40° field with a worst-case scalloping loss of 1.96 dB over 0.9–1.4 GHz (0.7-dB taper cost).

空间 FFT 波束形成与逐束直接求和最大归一差 4.1×10⁻⁵，运算量降低 8.6 倍Spatial-FFT beams match direct beam sums to 4.1×10⁻⁵ with 8.6 times less arithmetic SIM

数字链路Digital chain DESIGN

XCZU67DR RFSoC 采集板（每板 10 路 ADC）直接采样 L 波段、多相滤波信道化、重新量化为 4+4 位；百G交换网络送入 KU5P FPGA 卡做二维空间 FFT 成束；实时产品 2048 通道、约 0.1 ms，搜索后触发板载缓存（约 1.25 s）保存原始电压，服务器做相干消色散。XCZU67DR RFSoC acquisition boards (ten ADC inputs each) sample the L band directly, channelize with a polyphase filter bank and requantize to 4+4 bits; a 100G network feeds KU5P FPGA cards that form beams by 2-D spatial FFT. The real-time product has 2,048 channels at about 0.1 ms; a trigger freezes the on-board buffer (about 1.25 s) of raw voltages, which servers dedisperse coherently.

1000 万元档：280 块采集板、53 块成束卡、45 块搜索卡¥10 M tier: 280 acquisition boards, 53 beamforming and 45 search cards

分工审计Division of labour DATA

11 项规格层面的决策中 9 项由研究者做出，其余 2 项由研究者在 AI 给出的选项中选择或批准其提议；设计 agent 在 49 条人类消息之下进行了 2,729 次工具调用。没有任何决策由 AI 设定目标。The investigators made 9 of 11 specification-level decisions and, for the other 2, chose among options the AI offered or approved its proposal; the design agent made 2,729 tool calls for 49 human messages. No decision reached the level at which an AI sets the objective.

标签暂定，两名人类标注者正在重标Labels provisional; two human annotators are re-labelling PENDING

错误与交叉审阅Errors and cross-review DATA

规划 agent 的工作中共发现 44 处实质性错误：19 处在科学产出与成本模型，17 处在文字、图表与记录，8 处在另一项独立研究的方案中。跨模型审阅发现 25 处（含模型错误 19 处中的 14 处），红队 17 处，自查 2 处。44 substantive errors were found in the planning agent's work: 19 in the science-yield and cost models, 17 in text, figures and records, 8 in the protocol of a separate study. Cross-model review found 25 (including 14 of the 19 model errors), red-team review 17 and self-checks 2.

几处曾改变头条数字，例如遗漏脉冲占空比使首光时间从 2 分钟误算为 4 小时Several changed headline numbers; an omitted pulse duty cycle had put first light at 4 h instead of 2 min

三条设计规则Three design rules PRED

  • 2048 通道、0.1 ms 搜索加触发原始电压：约占预算 3%，51 µs 方案约 12%；等成本下产率相同。
  • 每路 400→700 MHz：超出模数转换器第一奈奎斯特区并碰到栅瓣和卫星业务，已撤回。
  • 两个外站约以一半探测数换取定位；需非共线布局与定标预算。
  • A 2,048-channel, 0.1-ms search with triggered raw voltages costs about 3% of the budget against 12% for 51 µs, with the same yield at equal cost.
  • Going from 400 to 700 MHz per input samples beyond the converters' first Nyquist zone and meets grating lobes and satellite services; withdrawn.
  • Two outriggers trade about half the detections for localization and need a non-collinear layout and a calibration budget.

前向模型（更完整的暴发族群模型）给出的单站亮 FRB 年数Bright FRBs per year for a single station from the forward models (fuller burst-population models) PRED

预算Budget经验模型：中位数（68% 区间）Empirical model: median (68% interval)宇宙学模型，全约束加权：中位数（68% 区间）Cosmological model, all-constraint weighting: median (68% interval)
1000 万元¥10 M17 (6–43)32 (15–54)
2000 万元¥20 M69 (29–144)92 (46–159)
3000 万元¥30 M154 (73–303)168 (87–291)
两个前向模型都以 ASKAP 复眼巡天定标，均未通过全部留出检验；区间主要反映亮端斜率的假设。本工具实时计算用的是更简单的规划模型（理想 1 ms 脉冲），其数值通常高于前向模型。Both forward models are calibrated on the ASKAP fly's-eye survey and neither passes every held-out test; the intervals mostly reflect the assumed bright-end slope. The live numbers in this tool come from the simpler planning model (ideal 1-ms pulse), which usually gives more.

模型与局限Model and limitations

怎么算的How it is computed

  • 按预算二分求出买得起的最大主站口径；单元数 = 填充系数 × 口径面积 / 阵距²，每单元两路输入。
  • 采集板按每板路数与 50G 链路容量推导；成束卡取“按输入带宽”和“按 FFT 运算量”两者的较大值；搜索卡按每卡波束数；交换端口含 1.5 倍汇聚冗余。
  • SEFD = 2kTsys/Ae；1 ms 注量阈值按两偏振 10σ、可用带宽 80%。
  • 亮 FRB 年数：ASKAP 全天率按亮端斜率外推到阈值，乘以视场内灵敏度加权的天区比例、365 天与 0.7 的占空。
  • 定位：每个方向至少两个伙伴站互相关信噪比达到 7。
  • The largest affordable main-station aperture is found by bisection; elements = packing × aperture area / pitch², with two inputs per element.
  • Acquisition boards follow from the inputs per board and the 50G link; beamforming cards are the larger of the input-bandwidth and FFT-arithmetic requirements; search cards follow from beams per card; switch ports include a 1.5-fold aggregation overhead.
  • SEFD = 2kTsys/Ae; the 1-ms fluence threshold assumes 10σ over two polarizations and 80% usable bandwidth.
  • Bright FRBs per year: the ASKAP all-sky rate scaled to the threshold with the bright-end slope, times the sensitivity-weighted sky fraction in the field, 365 days and a duty cycle of 0.7.
  • Localization: at least two partner stations must reach a cross-correlation S/N of 7.

推荐指向从哪来Where the pointing advice comes from

“按科学侧重推荐”使用前向模型在 1000 万元下对 76 种指向赤纬与视场组合的扫描结果；效用 = 磁星权重 × 磁星年数 + FRB 权重 × FRB 年数，各项先除以第0代（指向 30°、视场 40°）的值。磁星年数对 24 颗已确认磁星求和，以 SGR 1935+2154 的暴发定标，其均值由暴发率先验的上尾决定。The advice uses the forward-model scan of 76 combinations of pointing declination and field at ¥10 M; utility = magnetar weight × magnetar bursts + FRB weight × FRBs, each divided by its Gen-0 value (pointing 30°, field 40°). Magnetar bursts are summed over 24 confirmed magnetars and calibrated on SGR 1935+2154; their mean is set by the upper tail of the burst-rate prior.

视场超过 ±40° 的收益依赖把扫描损失外推到 40° 以外，尚未验证。Gains beyond ±40° rely on extrapolating scan losses beyond 40° and are not yet verified.

局限Limitations

  • 价格中服务器、宿主机箱、不可预见费为占位值；交换端口按二手价。
  • 尚未计入 FFT 网格到固定波束的插值、全交换转角、同步与定标硬件以及运行费用。
  • 规划模型不含脉宽分布、散射、通道内色散展宽与流程损失，产率偏乐观。
  • 竹节单元的电磁性能、模块灵敏度、站址干扰都待实测。
  • 所有数值都是设计、仿真或模型预测，不得当作测量引用。
  • Server, host-chassis and contingency prices are placeholders; switch ports use second-hand prices.
  • Not yet included: interpolation from the FFT grid to fixed beams, the all-to-all corner turn, synchronization and calibration hardware, and running costs.
  • The planning model omits width distributions, scattering, intra-channel smearing and pipeline losses, so its yields are optimistic.
  • The bamboo element's electromagnetic performance, module sensitivity and site interference all await measurement.
  • Every number is a design value, simulation or model prediction and must not be quoted as a measurement.

数字终端的两块板卡The two boards of the digital back end

整条数字链路只用两种商用板卡：XCZU67DR RFSoC 采集板直接采样并信道化，KU5P FPGA 卡负责波束合成和实时搜索。两者都以百G网络互连，可按台站规模线性增减。The whole digital chain uses only two commercial boards: the XCZU67DR RFSoC acquisition board samples and channelizes, and the KU5P FPGA card forms beams and runs the real-time search. Both connect over a 100G network and scale linearly with the size of the station.

XCZU67DR acquisition board (left) and KU5P FPGA card (right)
左：XCZU67DR RFSoC 采集板；右：KU5P FPGA 计算卡（不按比例）。照片由团队提供，背景经 ChatGPT（生成式 AI）修饰，板上细节以实物为准。Left: XCZU67DR RFSoC acquisition board; right: KU5P FPGA card (not to scale). Photographs supplied by the team; backgrounds retouched with ChatGPT (generative AI), so details on the boards may differ from the hardware.

XCZU67DR RFSoC 采集板（知芯电子）XCZU67DR RFSoC acquisition board (Zhixin Electronics) DESIGN

芯片DeviceAMD Zynq UltraScale+ RFSoC XCZU67DR，处理系统（PS）与可编程逻辑（PL）双域AMD Zynq UltraScale+ RFSoC XCZU67DR, with processing system (PS) and programmable logic (PL)
射频RF10 路 ADC 输入（2 路最高 5.9 GSPS，8 路最高 2.95 GSPS），8 路 DAC 输出（最高 10 GSPS）10 ADC inputs (2 up to 5.9 GSPS, 8 up to 2.95 GSPS) and 8 DAC outputs (up to 10 GSPS)
存储MemoryPS 侧 DDR4 4 GB（64 位），PL 侧 DDR4 2 GB（32 位）4 GB DDR4 on the PS (64-bit) and 2 GB on the PL (32-bit)
接口Interfaces2 个 QSFP（各 4 路高速串行）、千兆以太网、USB3、DisplayPort、JTAG、GPS 模块2 QSFP cages (4 serial lanes each), Gigabit Ethernet, USB3, DisplayPort, JTAG and a GPS module
时钟ClockingHMC7044 / SI53340 时钟链路，支持外部参考与同步输入输出HMC7044 / SI53340 clock chain with external reference and sync in/out
尺寸与供电Size and power178 mm × 130 mm, +12 V

在本设计中：In this design: 每板 10 路直接采样 L 波段（每路 400 MHz，1.0–1.4 GHz 落在 2.95 GSPS 的第一奈奎斯特区内），多相滤波信道化并重新量化为 4+4 位，以 50G 分支接入交换网络；DDR 中 5 GB 留作原始电压缓存，约 1.25 s。单价 1.5 万元，1000 万元档用 280 块，约占总成本 42%，是最大的成本项。each board samples 10 inputs directly in the L band (400 MHz per input; 1.0–1.4 GHz lies in the first Nyquist zone at 2.95 GSPS), channelizes with a polyphase filter bank, requantizes to 4 + 4 bits and connects to the switch through a 50G breakout; 5 GB of its DDR holds about 1.25 s of raw voltages. At ¥15,000 each, the ¥10 M tier uses 280 boards, about 42% of the cost and the largest single item.

KU5P FPGA 计算卡（微相 FNIC-KU5P）KU5P FPGA card (MicroPhase FNIC-KU5P) DESIGN

芯片DeviceAMD Kintex UltraScale+ XCKU5P，1,824 个 DSP48E2（分销商资料，待与 AMD 数据手册核对）AMD Kintex UltraScale+ XCKU5P with 1,824 DSP48E2 slices (distributor data, to be checked against the AMD data sheet)
网络Network前面板 2 个 QSFP28，各支持 100GbE2 QSFP28 ports on the front panel, 100GbE each
形态Form factorPCIe 卡，由宿主机箱的 PCIe 槽供电；本设计按每台机箱 8 块计PCIe card powered by the host's PCIe slot; this design assumes 8 cards per host chassis
单价Price1,700 元（研究者提供）¥1,700 (supplied by the investigators)

在本设计中：In this design: 同一种板卡承担两项任务。成束卡对网格化的口径电压做二维空间 FFT，形成覆盖 ±40° 的固定波束，卡数取“按输入带宽”和“按 DSP 算力”两者的较大值，1000 万元档 53 块；搜索卡做消色散与分类，每卡处理 25 个 2048 通道、约 0.1 ms 的波束，1000 万元档 45 块。两者合计约占成本 2%。the same card does two jobs. Beamforming cards apply a 2-D spatial FFT to the gridded aperture voltages to form fixed beams over ±40°; their number is the larger of the input-bandwidth and DSP-arithmetic requirements, 53 at the ¥10 M tier. Search cards dedisperse and classify, 25 beams of 2,048 channels at about 0.1 ms per card, 45 at the ¥10 M tier. Together they are about 2% of the cost.

成束卡数尚未计入 FFT 网格到固定波束的插值（可能增加数倍）；定点系数、存储带宽与时序待 FPGA 综合验证The beamforming count omits the interpolation from the FFT grid to fixed beams (which may multiply it); fixed-point coefficients, memory bandwidth and timing await FPGA synthesis PENDING