lensglobal op• 数据种类:table → table
• 调用: import fullseye as fs; fs.ledger.lens_global_search(system=None, variables=None, merit=None, n_starts: 'int' = 8, seed: 'int' = 0, spread: 'float' = 0.5, iterations: 'int' = 40, duplicate_tolerance: 'float' = 0.001) -> 'dict'(要直接调用实现,import lensglobal; lensglobal.lens_global_search(system=None, variables=None, merit=None, n_starts: 'int' = 8, seed: 'int' = 0, spread: 'float' = 0.5, iterations: 'int' = 40, duplicate_tolerance: 'float' = 0.001) -> 'dict';从台账取用则 opsoptics.get("lens_global_search"))
用多起点局部优化(lensopt.optimize_lens)收集局部最优解,按 merit 顺序返回(table)。
> 以下的详细说明为原文 —— 摘要与标题已翻译。
Start 0 is *system* as given; starts 1 .. *n_starts*-1 move every variable to x (1 + spread u) with u uniform in [-1, 1] drawn from numpy.random.default_rng(seed) (a variable that is zero, such as a flat, moves by *spread* times the typical size of its kind). A perturbed start that is not a valid prescription is redrawn (up to 25 times, then ValueError). Every start is optimised by lensopt.optimize_lens with the *same* merit and the *same* EFL target (the EFL of *system* unless merit["efl_target"] says otherwise), so the optima are comparable. *merit* is a dict of fields wavelengths rings efl_target efl_weight field_weights pupil_fill min_thickness max_thickness min_radius damping tolerance; *variables* are as in optimize_lens.
Returns {"optima", "best", "n_starts", "n_distinct", "basins", "seed", "spread"}. optima holds every start's result sorted by merit (ties by start index): {"start", "merit", "start_merit", "rms_spot", "efl", "system", "values", "status", "converged", "duplicate_of"} where duplicate_of is the start of an earlier (better) optimum with the same variables (every variable within *duplicate_tolerance* relative, floor 1e-9) or None. basins lists the distinct optima with the number of starts that reached each. Deterministic for a given seed.
Example::
r = lens_global_search(singlet, ["c0", "c1"], {"efl_target": 100.0}, n_starts=12, seed=3)
r["best"]["merit"], r["n_distinct"]
optics 的每个算子都先校验输入再计算(不让任何东西无声通过):
• 单位写进参数名 —— _mm / _um / _deg / _mrad。把 mm 和 µm 弄混不会崩溃,而是给出「看着合理却是错的答案」,所以用命名来防。这里绝不从数值大小去猜单位。
• 字符串一律 ValueError —— float('50') 会成功,于是未解析的配置值会被当成长度混进来(实测:thin_lens('50', '200') 曾返回看着合理的 66.667 mm)。bool 也按 True == 1 的隐式提升拒绝。
• complex / masked array 一律 ValueError(仅接受实数槽位;拒绝无声丢弃虚部或剥掉掩码)。所有输入中的 NaN/Inf 一律 ValueError。
• 逐项点名拒绝除零及其近亲:焦距 0、曲率半径 0、折射率 <= 0、全不透明光阑(全为 0,归一化变成 0/0)、总和 <= 0 的 PSF、S0 = 0 的 Stokes 矢量、物体位于前焦点(像在无穷远)。
• 只有两个算子会返回非有限值,而且都写进了契约:depth_of_field 在超焦距以外返回 far_mm = inf(这正是超焦距的定义),gaussian_beam 在束腰处返回 wavefront_radius_mm = inf(平面波前的曲率半径)。两者都同时返回一个有限的搭档(far_is_infinite / curvature_per_mm)。除此之外的无声 NaN/Inf 都在内部检出并 ValueError ——「float64 溢出了」和「答案是无穷大」是两种不同的主张,不能拿后者的脸去交付前者。
• 尺寸上限:生成网格受 optics.MAX_GRID(4096)限制,传入的场/PSF/光阑受 optics.MAX_FIELD_ELEMENTS(2^24),ABCD 元件序列受 optics.MAX_SYSTEM_ELEMENTS(1024),Zernike 受 MAX_ZERNIKE_TERMS(512)/ MAX_ZERNIKE_ORDER(40)/ MAX_ZERNIKE_BASIS(2^25)。以 fail-closed 堵住「小参数引发巨大内部分配」的路径(实测:n_max=40 × 4096² 需要 108 GB)。
• 物理上不可能的状态同样拒绝:偏振度 > 1 的 Stokes 矢量、负透过率、负强度、n-|m| 为奇数等非法 Zernike 指标。
• 示例数据目录(下载 URL / 许可证) —— 2-D 用 skimage.data(BSD/公有领域)加合成图,3-D 给出真实数据源(Stanford/PDS 等)的下载 URL。
• 算子来历与参考文献 —— 该算子族所依据的研究/方法出处。
• 算法的正典(作者、年份)与用途见上面的族使用指南。
• poc_lens_global_search — py -3.11 examples/poc_lens_global_search.py
table 作为输入)abcd_matrix · wavefront_stats · sfr_from_edge · paraxial_trace · seidel_coefficients · spot_stats · tolerance_analysis · wavefront_from_opd
lensglobal)glass_nearest · lens_substitute_glasses
*Provenance: lensglobal.py — OPTICS 算子登记表。本条目由 tools/opdocs.py md 自动生成(请勿手工编辑)。*
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.