Files
Jungfraujoch/image_analysis/rotation_indexer/RotationIndexer.cpp
T
leonarski_fandClaude Opus 5 0fccbe21b5
Build Packages / Create release (push) Successful in 15s
Build Packages / build:rugnux:aarch64 (cross) (push) Successful in 7m21s
Build Packages / build:rugnux-tgz (x86_64) (push) Successful in 8m28s
Build Packages / build:viewer-tgz:cpu (push) Successful in 9m53s
Build Packages / build:viewer-tgz:cuda (push) Successful in 11m57s
Build Packages / build:rpm (rocky8_nocuda) (push) Successful in 15m12s
Build Packages / build:rpm (rocky9_nocuda) (push) Successful in 15m50s
Build Packages / build:windows:nocuda (push) Successful in 17m15s
Build Packages / build:windows:cuda (push) Successful in 19m50s
Build Packages / HDF5 consumer tests (DIALS, XDS) (push) Successful in 23m14s
Build Packages / build:rpm (ubuntu2204_nocuda) (push) Successful in 19m9s
Build Packages / build:rpm (ubuntu2404_nocuda) (push) Successful in 19m12s
Build Packages / build:rugnux:windows (push) Successful in 10m50s
Build Packages / Generate python client (push) Successful in 44s
Build Packages / Build documentation (push) Successful in 1m23s
Build Packages / build:rpm (rocky8_sls9) (push) Successful in 20m10s
Build Packages / build:rpm (rocky9_sls9) (push) Successful in 19m23s
Build Packages / build:rpm (rocky8) (push) Successful in 18m2s
Build Packages / build:rpm (rocky9) (push) Successful in 18m22s
Build Packages / build:rpm (ubuntu2204) (push) Successful in 13m46s
Build Packages / build:rpm (ubuntu2404) (push) Successful in 12m25s
Build Packages / Unit tests (push) Successful in 1h12m4s
symmetry: the cell a group is adopted on is refined under that group
A space group confirmed from the intensities AFTER integration is one no
constrained fit has produced. The Bravais class is decided on the unrefined
indexing candidate, so a two-fold the spot positions never offered leaves the
freely refined metric standing in the group's own setting - a C 1 2 1 whose
alpha is 88.5 - and that cell goes to the report, the master file and the MTZ.

At adoption, re-refine the lattice under the group's constraint against the
accumulated rotation spots, at the geometry the images were integrated at, and
keep it when the spots do - the bar the indexer's own pseudo-symmetry guard
uses. Where they refuse it, report the nearest metric the group fixes and say
that a deviation that size is not refinement noise. A cell whose violation is
above MAX_METRIC_VIOLATION is the WRONG cell for its group rather than an
unconstrained one, and nothing here touches it: a visible mismatch must not
become a plausible-looking one.

The projection is applied as a change of basis, not as three rebuilt vectors:
the three-Coord constructor enforces a right-handed basis, and a left-handed
lattice came back with an axis flipped and its free angle replaced by the
supplement, which failed the merge outright.

Battery over 151 datasets, base against this: 144 byte-identical, 7 cells moved
onto their group's metric, no space group changed, no run gained or lost, open
arm 94/99 both ways. Every merge statistic of the seven is unchanged except
completeness, which rises on four and falls 0.1 % on one.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KWkZ1o2aoQ9EimF2wtzBky
2026-09-15 15:57:12 +02:00

682 lines
38 KiB
C++

// SPDX-FileCopyrightText: 2025 Filip Leonarski, Paul Scherrer Institute <filip.leonarski@psi.ch>
// SPDX-License-Identifier: GPL-3.0-only
#include "../../common/JFJochMath.h"
#include "RotationIndexer.h"
#include "../geom_refinement/XtalOptimizer.h"
#include "../indexing/FFTIndexer.h"
#include "../lattice_search/LatticeSearch.h"
#include "../indexing/MultiLatticeSearch.h"
#include <future>
namespace {
// Sub-cell override thresholds used in candidate selection to undo a spurious axis doubling:
// a later candidate replaces the chosen cell when it is smaller by more than this volume ratio
// (a doubling is 2x, well past 1.5) and indexes within this fraction slack of it. The slack is
// far below the indexed-fraction gap a real superstructure opens between its true cell and its
// sub-cell, so genuine large cells are kept.
constexpr float ROT_SUBCELL_VOLUME_RATIO = 1.5f;
constexpr float ROT_SUBCELL_FRAC_SLACK = 0.02f;
// How much better a lower-symmetry SETTING of an already-chosen lattice has to index before it is
// taken (see the selection below). A subgroup setting holds fewer cell parameters fixed, so it can
// never index less - only a decisively better fit is evidence that the higher symmetry is wrong.
constexpr float ROT_SUBGROUP_FRAC_RATIO = 1.5f;
// The same, for a candidate whose primitive cell is a near-integer MULTIPLE of the chosen one.
// Multiplying an axis halves that reciprocal spacing, so the multiple has a lattice point
// wherever its sub-cell has one and another in between: it collects spots the sub-cell leaves
// unindexed, for reasons that have nothing to do with the crystal. The indexed fraction is
// biased in its favour, so a small lead over the sub-cell is not evidence, and deciding the
// pair on the ordinary margin leaves it turning on the last bits of that fraction - one
// crystal here separated the true cell from a spurious 5x supercell by 0.003, little enough
// that the build's -march flags settled it, and the supercell merged to an R-free of 0.58
// (i.e. noise). A real superstructure's satellite rows are a large share of its spots and
// clear this ratio comfortably - but see the limitation noted at the comparison itself.
constexpr float ROT_SUPERCELL_FRAC_RATIO = 1.5f;
// How far from a whole number the volume ratio may sit and still count as an axis multiple.
constexpr double ROT_SUPERCELL_INTEGER_TOL = 0.15;
// Iteration bound for the candidate-cell refinements when the run is not real-time. Without one,
// XtalOptimizerData falls back to its wall-clock bound and the cell a candidate refines to - and
// therefore which candidate wins - depends on how busy the machine was. Same value as Ceres' own
// default iteration limit, so it only bites where the clock was biting before.
constexpr int ROT_REFINE_ITERATIONS = 50;
// Order of the lattice point group, so "lower symmetry" is a well-defined comparison
// (gemmi's enum orders Trigonal after Tetragonal, which have 6 and 8 rotations).
int LatticePointGroupOrder(gemmi::CrystalSystem s) {
switch (s) {
case gemmi::CrystalSystem::Monoclinic: return 2;
case gemmi::CrystalSystem::Orthorhombic: return 4;
case gemmi::CrystalSystem::Trigonal: return 6;
case gemmi::CrystalSystem::Tetragonal: return 8;
case gemmi::CrystalSystem::Hexagonal: return 12;
case gemmi::CrystalSystem::Cubic: return 24;
default: return 1; // Triclinic
}
}
// Re-express a primitive hexagonal/trigonal lattice in the conventional hexagonal setting
// (a = b, gamma = 120). The Niggli-reduced primitive cell carries the two equal-length axes
// at gamma = 60; replacing b with b - a opens that angle to 120 without changing the lattice.
CrystalLattice HexagonalConventional(CrystalLattice latt) {
latt.ReorderABEqual(); // put the equal-length pair in a, b
Coord a = latt.Vec0(), b = latt.Vec1(), c = latt.Vec2();
if (angle_deg(a, b) < 90.0f)
b -= a;
return CrystalLattice(a, b, c); // constructor fixes handedness
}
bool IsHexagonalSystem(gemmi::CrystalSystem s) {
return s == gemmi::CrystalSystem::Trigonal || s == gemmi::CrystalSystem::Hexagonal;
}
// The hexagonal lattice metric (two equal axes at 60/120 deg, both perpendicular to the third) is
// also satisfied by its ortho-hexagonal C-centred supercell, so the geometry-keyed LatticeSearch can
// land there. Detect the hexagonal metric on the reduced PRIMITIVE cell so the de-novo path (no space
// group to key on) can re-express it in conventional hexagonal axes.
bool IsMetricallyHexagonal(CrystalLattice latt, float rel_tol = 0.03f, float angle_tol_deg = 3.0f) {
latt.ReorderABEqual();
const Coord a = latt.Vec0(), b = latt.Vec1(), c = latt.Vec2();
const float la = a.Length(), lb = b.Length();
if (la <= 0.0f || lb <= 0.0f || std::fabs(la - lb) > rel_tol * std::max(la, lb))
return false;
const float gab = angle_deg(a, b);
if (std::fabs(gab - 60.0f) > angle_tol_deg && std::fabs(gab - 120.0f) > angle_tol_deg)
return false;
return std::fabs(angle_deg(a, c) - 90.0f) <= angle_tol_deg &&
std::fabs(angle_deg(b, c) - 90.0f) <= angle_tol_deg;
}
// Fraction of the accumulated reciprocal-space spots that a lattice indexes to near-integer
// Miller indices within tol. Comparing a symmetry-constrained refinement against an
// unconstrained (triclinic) one is a data-driven test for a false promotion: a wrong
// higher-symmetry constraint snaps a pseudo cell onto ideal angles and misplaces most spots.
float IndexedFraction(const CrystalLattice &latt, const std::vector<Coord> &coords, float tol) {
if (coords.empty())
return 0.0f;
const Coord a = latt.Vec0(), b = latt.Vec1(), c = latt.Vec2();
const float tol_sq = tol * tol;
size_t indexed = 0;
for (const Coord &s : coords) {
const float dh = a * s - std::round(a * s); // Coord operator* = dot product = Miller index
const float dk = b * s - std::round(b * s);
const float dl = c * s - std::round(c * s);
if (dh * dh + dk * dk + dl * dl < tol_sq)
++indexed;
}
return static_cast<float>(indexed) / static_cast<float>(coords.size());
}
}
RotationIndexer::RotationIndexer(const DiffractionExperiment &x, IndexerThreadPool &indexer,
bool real_time)
: experiment(x),
index_ice_rings(x.GetIndexingSettings().GetIndexIceRings()),
real_time(real_time),
v_(experiment.GetImageNum()),
angle_deg_(experiment.GetImageNum()),
axis_(x.GetGoniometer()),
geom_(x.GetDiffractionGeometry()),
updated_geom_(geom_),
indexer_(indexer) {
}
void RotationIndexer::RunIndexing() {
std::unique_lock ul(m);
if (!axis_)
return;
std::vector<Coord> coords;
coords.reserve(accumulated_spots); // the exact count held; the per-image cap may be SIZE_MAX
for (int i = 0; i < v_.size(); i++) {
const float angle_deg = angle_deg_[i].value_or(axis_->GetAngle_deg(i) + axis_->GetWedge_deg() / 2.0f);
const auto rot = axis_->GetTransformationAngle(angle_deg);
for (const auto &s: v_[i])
coords.emplace_back(rot * s.ReciprocalCoord(geom_));
}
const auto indexer_result = indexer_.Run(experiment, coords);
indexer_error_ = indexer_result.error;
if (!indexer_result.lattice.empty() && indexer_result.lattice[0].CalcVolume() > 1.0) {
DiffractionExperiment experiment_copy(experiment);
const float index_tol = experiment.GetIndexingSettings().GetTolerance();
const auto orig_axis = axis_;
// Map an FFT candidate cell to its Bravais lattice. Re-express a metrically-hexagonal cell in
// conventional hexagonal axes (LatticeSearch can land on the ortho-hexagonal C setting) so the
// 3-fold is not hidden from scaling.
//
// A user-fixed space group is deliberately NOT stamped on here. The group names the symmetry;
// it does not say which basis the FFT candidate came back in, and the conventional cell above
// was reduced for whatever class the METRIC matched. Relabelling that cell with the group's
// system and centring leaves the constrained refine snapping the wrong angles to the ideal
// ones: measured, a C-centred orthorhombic cell relabelled primitive monoclinic indexed 1 of
// 60 validation frames and an F-cubic one relabelled trigonal indexed 0 of 60, where the same
// frames index at 36/60 and 51/60 without a group. The group is applied where it belongs - to
// the scaling and the merge.
auto build_sr = [&](const CrystalLattice &cand) -> LatticeSearchResult {
auto ls = LatticeSearch(cand);
if (!IsHexagonalSystem(ls.system) && IsMetricallyHexagonal(ls.primitive_reduced)) {
ls.conventional = HexagonalConventional(ls.primitive_reduced);
ls.system = gemmi::CrystalSystem::Hexagonal;
ls.centering = 'P';
}
return ls;
};
// Re-accumulate the reciprocal spots under a refined geometry/axis, to score a refined cell.
auto accumulate = [&](const DiffractionGeometry &g, const std::optional<GoniometerAxis> &ax) {
return AccumulateReciprocal(g, ax);
};
// The FFT offers a few candidate cells (its best reduction plus, for large/elongated cells, a
// widened alternative). Fully refine each and keep the one that indexes the most spots AFTER
// geometry refinement - the pre-refinement fraction is not a reliable discriminator (an
// incorrect larger cell can fit more of the un-refined accumulated spots than the correct one).
// Twelve, not four. The fraction that orders the candidates is the unreliable one named
// above, so the correct cell is not always in the first few: on a crystal whose shortlist
// carries near-degenerate reductions, three of the first four slots went to cells that
// cannot exist. Refining a candidate is cheap next to the pass that produced it.
const size_t n_try = std::min<size_t>(indexer_result.lattice.size(), 12);
// Bound the axis lengths just above the found cell so a free (triclinic) refine cannot drift
// onto a pseudo-translation / modulation supercell (a modulated crystal whose satellites
// define a ~4x period would otherwise inflate one axis to the max-length clamp).
auto make_data = [&](const CrystalLattice &latt, gemmi::CrystalSystem sys, float length_bound_A) {
XtalOptimizerData d{
.geom = experiment_copy.GetDiffractionGeometry(),
.latt = latt,
.crystal_system = sys,
.min_spots = experiment.GetIndexingSettings().GetViableCellMinSpots(),
.max_length_A = length_bound_A,
// Match the indexers' [30,150] deg bound so a monoclinic beta outside [60,120]
// (e.g. beta>120) is refined, not clamped to the boundary.
.min_angle_deg = 30.0f,
.max_angle_deg = 150.0f,
.refine_beam_center = true,
.refine_detector_angles = true,
.refine_rotation_axis = true,
.index_ice_rings = experiment.GetIndexingSettings().GetIndexIceRings(),
.max_iterations = real_time ? 0 : ROT_REFINE_ITERATIONS,
.axis = orig_axis
};
if (d.crystal_system == gemmi::CrystalSystem::Trigonal)
d.crystal_system = gemmi::CrystalSystem::Hexagonal;
if (d.crystal_system == gemmi::CrystalSystem::Monoclinic)
d.latt.ReorderMonoclinic();
return d;
};
// Refine the FFT candidates. Each candidate is independent, and within a candidate the
// metric-symmetry solve and the de-novo triclinic pseudo-symmetry solve are independent too,
// so refine all of them (up to ~8 solves) at once - these Ceres refinements are the dominant
// first-pass cost. Each solve runs Ceres on a few cores. Selection stays serial and in
// candidate order below, so the outcome is identical to refining them one by one.
constexpr int kCeresThreads = 4;
// Seed each candidate serially (cheap: LatticeSearch + setup), then solve them in parallel.
struct CandidateWork {
bool viable = false;
LatticeSearchResult sr;
XtalOptimizerData constrained;
bool has_tri = false;
XtalOptimizerData tri;
};
std::vector<CandidateWork> work(n_try);
for (size_t ci = 0; ci < n_try; ci++) {
const CrystalLattice &cand = indexer_result.lattice[ci];
if (cand.CalcVolume() <= 1.0)
continue;
CandidateWork &w = work[ci];
w.sr = build_sr(cand);
const auto conv_uc = w.sr.conventional.GetUnitCell();
const float length_bound_A = 1.2f * static_cast<float>(std::max({conv_uc.a, conv_uc.b, conv_uc.c}));
w.constrained = make_data(w.sr.conventional, w.sr.system, length_bound_A);
// Pseudo-symmetry guard: also refine unconstrained (triclinic) on the primitive cell. Run it
// with a user-fixed space group too. The metric promotion the constrained refine acts on is
// decided by the geometry, not by the group, so it can be false whether or not a group was
// given - and without this cell there is nothing to catch it with, which is how a fixed
// group turned crystals the de-novo path indexes at 60/60 into runs that index none.
w.has_tri = (w.sr.system != gemmi::CrystalSystem::Triclinic);
if (w.has_tri)
w.tri = make_data(w.sr.primitive_reduced, gemmi::CrystalSystem::Triclinic, length_bound_A);
w.viable = true;
}
// Refine (constrained metric solve + score by the refined-geometry indexed fraction, the
// reliable discriminator). Runs on its own thread per solve.
struct Solved { bool ok = false; float frac = 0.0f; XtalOptimizerData data; };
auto solve = [&](XtalOptimizerData d) -> Solved {
const bool ok = XtalOptimizer(d, v_, kCeresThreads);
const float frac = ok ? IndexedFraction(d.latt, accumulate(d.geom, d.axis), index_tol) : 0.0f;
return {ok, frac, std::move(d)};
};
std::vector<std::future<Solved>> constrained_f(n_try), tri_f(n_try);
for (size_t ci = 0; ci < n_try; ci++) {
if (!work[ci].viable)
continue;
constrained_f[ci] = std::async(std::launch::async, [&, ci] { return solve(work[ci].constrained); });
if (work[ci].has_tri)
tri_f[ci] = std::async(std::launch::async, [&, ci] { return solve(work[ci].tri); });
}
// Assemble and select serially, in candidate order - identical to refining them one by one.
float best_frac = -1.0f;
float best_prim_vol = 0.0f;
bool have_best = false;
size_t best_ci = 0;
XtalOptimizerData best_data;
LatticeSearchResult best_sr;
std::shared_ptr<RotationIndexerResult> best_alt;
for (size_t ci = 0; ci < n_try; ci++) {
if (!work[ci].viable)
continue;
Solved c = constrained_f[ci].get();
bool ok = c.ok;
float frac = c.frac;
XtalOptimizerData data = std::move(c.data);
LatticeSearchResult sr = work[ci].sr;
// Adopt the free triclinic cell only if it indexes CLEARLY more than the constrained cell -
// a false promotion (a near-90 pseudo cell forced to ideal angles + a bogus centering)
// misplaces most reflections (measured indexed-fraction ratio ~0.1), whereas genuine higher
// symmetry (incl. R-centred) indexes comparably (ratio ~0.7). Preferring the constrained
// cell on a near-tie keeps the real symmetry/centering; the intensities settle the final
// space group.
// When the guard leaves the constrained cell in place, hand the refined triclinic cell to
// the caller instead of dropping it: the accumulated-spot fraction separates a false
// promotion from genuine symmetry by less than 2x on a lattice that is pseudo-symmetric
// to a few tenths of a degree, while the caller's per-frame validation separates the same
// pair by more than 20x.
std::shared_ptr<RotationIndexerResult> tri_alt;
if (work[ci].has_tri) {
Solved t = tri_f[ci].get();
auto as_triclinic = [](LatticeSearchResult s) {
s.system = gemmi::CrystalSystem::Triclinic;
s.centering = 'P';
s.conventional = s.primitive_reduced;
s.reindex = gemmi::Mat33(1, 0, 0, 0, 1, 0, 0, 0, 1);
return s;
};
if (t.ok && t.frac > 0.3f && frac < 0.5f * t.frac) {
sr = as_triclinic(sr);
data = std::move(t.data);
ok = true;
frac = t.frac;
} else if (t.ok) {
tri_alt = std::make_shared<RotationIndexerResult>(RotationIndexerResult{
.lattice = t.data.latt,
.search_result = as_triclinic(sr),
.geom = t.data.geom,
.axis = t.data.axis,
});
}
}
if (!ok)
continue;
// Prefer the indexer's earlier (primary) candidate; adopt a later one only if it indexes
// clearly more AND indexes reasonably well in absolute terms. The absolute floor stops a
// marginally-higher alternative from displacing the primary when both index poorly (e.g. a
// twin, where the accumulated-spot fraction is a noisy proxy) - only a decisively better
// cell (a superstructure's true cell vs its sublattice) takes over.
// Displace the current best when the candidate indexes clearly more, OR when it is a
// genuine sub-cell: a meaningfully smaller cell that still indexes at least as many spots.
// The sub-cell branch unmasks a spurious supercell (axis doubling): the primitive cell
// always indexes >= its integer multiple, so a doubled cell that wins ci-order by the
// hysteresis margin is overridden by its own primitive. A real superstructure's true
// (larger) cell indexes MORE than its sub-cell and is kept by the clearly-more branch;
// twin lattices share the cell volume, so this never disturbs twin selection.
// Volumes are compared PRIMITIVE. A centred conventional cell is an exact integer multiple
// of its primitive one, so two settings of the same lattice differ by that factor and
// conventional volumes read a mere change of setting as a sub-cell. And two such settings
// are not comparable on the indexed fraction either: the lower-symmetry one holds fewer
// cell parameters fixed, so it can only index more. An F-cubic lattice contains an
// I-tetragonal cell of the same volume, and refining that cell frees the c/a ratio the
// cubic one holds at sqrt(2), buying back the spots a fraction of a percent of strain had
// put out of tolerance. A slightly higher fraction is therefore no evidence against the
// higher symmetry - only a decisively better fit is.
const float cand_prim_vol = std::abs(data.latt.ToPrimitive(sr.centering).CalcVolume());
const bool lower_symmetry_setting = have_best
&& LatticePointGroupOrder(sr.system) < LatticePointGroupOrder(best_sr.system)
&& std::abs(cand_prim_vol - best_prim_vol) < 0.05f * best_prim_vol;
// A near-integer volume multiple of the incumbent - see ROT_SUPERCELL_FRAC_RATIO.
const double vol_ratio = (have_best && best_prim_vol > 1.0f && cand_prim_vol > 1.0f)
? cand_prim_vol / best_prim_vol : 1.0;
const double vol_nearest = std::round(vol_ratio);
const bool integer_supercell = vol_nearest >= 2.0
&& std::abs(vol_ratio - vol_nearest) < ROT_SUPERCELL_INTEGER_TOL;
// NOTE, and it is a real limitation: `frac > RATIO * best_frac` cannot be satisfied at
// all once best_frac exceeds 1/RATIO - above 0.667 for a ratio of 1.5, which is ordinary
// for good rotation data. So on data that indexes well these two guards do not merely
// raise the bar, they close the branch: no axis multiple and no lower-symmetry setting
// can displace the incumbent however much better it fits. A genuine superstructure whose
// satellite rows the sub-cell misses is therefore kept as its sub-cell, silently.
//
// Restating the bar on the fraction left UNINDEXED - the candidate must account for
// 1/RATIO of what the incumbent missed - is well defined over the whole range and looks
// like the obvious repair. It was tried and it REGRESSED the 37-crystal battery from
// 34/37 to 32/37 correct space groups: one C2 lattice fell to P1, and a P2 case went to
// C222 keeping 2923 of 22440 reflections. The indexed fraction is too noisy a statistic
// to carry a looser test, so the unreachable-but-safe form stays until the selection is
// decided on something better than it.
const bool clearly_more = frac > best_frac + 0.05f && frac > 0.15f
&& (!lower_symmetry_setting || frac > ROT_SUBGROUP_FRAC_RATIO * best_frac)
&& (!integer_supercell || frac > ROT_SUPERCELL_FRAC_RATIO * best_frac);
const bool smaller_subcell = have_best && frac > 0.15f
&& frac >= best_frac - ROT_SUBCELL_FRAC_SLACK
&& cand_prim_vol < best_prim_vol / ROT_SUBCELL_VOLUME_RATIO;
if (!have_best || clearly_more || smaller_subcell) {
best_frac = frac;
best_prim_vol = cand_prim_vol;
have_best = true;
best_data = std::move(data);
best_sr = sr;
best_ci = ci;
best_alt = std::move(tri_alt);
}
}
// Drive the selected candidate to the fit's FIXED POINT. Refinement here is a chain - solve,
// re-accumulate the reciprocal-space cloud under the refined geometry, solve again - and the
// selection above calls the solver ONCE per candidate, so what it returns is a point on the way
// to that fixed point rather than the fixed point itself. One round is enough to rank the
// candidates; it is not enough to have measured the geometry.
//
// What that costs, on a sweep whose 2theta reaches far enough for the detector tilt to be
// determined at all (long wavelength, short distance): the spindle-parallel tilt comes out at
// 16 % of an independently measured value after one round and 87 % of it after twenty, and the
// three sweeps of one crystal at three wavelengths then agree with that external value to
// 0.02 deg. The data follow - R_meas 0.137 -> 0.119, ISa 7.8 -> 9.2, and the anomalous peak
// height at the sulphur positions of a known model 7.0 -> 7.8 sigma.
//
// It also removes the fit's dependence on where the beam centre started, which is the thing
// the gauge prior below exists to guard: displacing the starting centre over 8 px moves the
// refined tilt by 0.0004 deg/px here against 0.0077 deg/px for a single round, on a geometric
// one-for-one of 0.0710. So the prior stays, and iterating turns it from a prior that pins the
// answer near the header into a per-step limit that lets the pair walk to their joint optimum -
// pinning its anchor across the rounds instead leaves the tilt at 0.014 deg and the data worse
// than doing nothing. The winner only: the other candidates are discarded, and these solves are
// the dominant first-pass cost. Not in real time, where the budget is wall-clock.
//
// The chain is a trajectory, and its last point is not always its best one. Every solve ends
// by fitting only the spots inside its tightest gate (0.1), so a cell with a direction the
// data barely speak about - which is what a free cell whose metric is near a Bravais class
// has - can slide along that direction, pulling a core of spots tighter while the periphery
// falls out of the fit altogether. Measured on such a crystal: over seventeen rounds the
// spots inside the tight gate rise from 0.295 to 0.327 while the spots inside the wide one
// peak at round three (0.754) and fall to 0.730, and it is round three that merges - ISa 11.0
// against 6.3, R_meas 0.148 against 0.213 - and that agrees with the archived reference cell.
// Which round a build stopped on used to decide that, since the tilt step the loop tests is
// scatter at the size of its own bound.
//
// So score every round on the WIDE gate the last pass does not fit - the same 0.3 the solve's
// first pass selects on - and commit the best-scoring round rather than the one the loop
// happens to stop on. The comparison is between rounds of ONE chain, whose cell moves by a
// fraction of a percent, so it is not a test a bigger cell can win by being bigger.
//
// The score is a COUNT of spots, and a lead of fewer than sqrt(count) of them is smaller than
// the count's own noise - taking the round on one would hand the answer back to the last
// bits, which is the defect being removed. So the chain's last round stands unless another
// beats it by more than that, and the ordinary chain - which settles, and whose rounds then
// score within a spot or two of each other - commits exactly what it committed before. The
// bound is generous: both counts are made over the same spot list, so the difference of two
// rounds carries far less noise than sqrt of either.
//
// And the round taken has to be the less distorted lattice as well as the better-fitting one.
// Ask LatticeSearch for the class of each round's cell - to MEASURE the drift, not to impose
// the class, which is measured fatal - and take the earlier round only when it matched the
// SAME class and its metric sits closer to it. Same class, because the deviation is a
// fraction of the tolerance of whatever class the round matched, so two classes' deviations
// are not the same quantity - and a round that matched no class reports 0, which is not
// "undistorted" but "nothing was asserted". Comparing that against a class's deviation reads
// the absence of a constraint as the absence of distortion, and the veto silently switches
// off. Measured: it would do that on 11 of 1132 traced chains, four of them against a
// round that matched no class at all.
//
// A lattice's metric symmetry is exact, so among cells that all fit the data the least
// distorted one is the one the crystal has; a chain that is merely converging does not move
// its distortion (a symmetry-constrained solve holds it at zero throughout), so this fires on
// exactly the chains that are sliding out of a class and on no others. It is a veto and not
// the criterion: the spots have to prefer the round first, so this cannot pull an answer
// towards a symmetry the data do not support.
constexpr int ROT_REFINE_OUTER_ROUNDS = 20;
constexpr double ROT_REFINE_TILT_SETTLED_RAD = 1.0e-5; // ~0.6 mdeg
if (have_best && !real_time) {
auto wide_count = [&](const XtalOptimizerData &d) {
const auto c = accumulate(d.geom, d.axis);
return IndexedFraction(d.latt, c, XTAL_OPTIMIZER_WIDE_TOLERANCE) * static_cast<float>(c.size());
};
// True when a matched the same Bravais class as b and sits closer to its ideal metric.
auto less_distorted = [](const LatticeSearchResult &a, const LatticeSearchResult &b) {
return a.system == b.system && a.centering == b.centering
&& MetricDeviation(a) < MetricDeviation(b);
};
XtalOptimizerData kept = best_data;
float kept_count = -1.0f;
float last_count = 0.0f;
LatticeSearchResult kept_class, last_class;
for (int r = 0; r < ROT_REFINE_OUTER_ROUNDS; ++r) {
XtalOptimizerData d = best_data;
if (!XtalOptimizer(d, v_, kCeresThreads))
break;
const double d1 = std::abs(d.geom.GetPoniRot1_rad() - best_data.geom.GetPoniRot1_rad());
const double d2 = std::abs(d.geom.GetPoniRot2_rad() - best_data.geom.GetPoniRot2_rad());
best_data = std::move(d);
last_count = wide_count(best_data);
last_class = LatticeSearch(best_data.latt);
if (last_count >= kept_count) {
kept_count = last_count;
kept_class = last_class;
kept = best_data;
}
if (std::max(d1, d2) < ROT_REFINE_TILT_SETTLED_RAD)
break;
}
if (kept_count > last_count + std::sqrt(last_count) && less_distorted(kept_class, last_class))
best_data = std::move(kept);
best_frac = IndexedFraction(best_data.latt, accumulate(best_data.geom, best_data.axis), index_tol);
}
if (have_best) {
search_result_ = best_sr;
indexed_lattice = best_data.latt;
updated_geom_ = best_data.geom;
axis_ = best_data.axis;
unconstrained_ = std::move(best_alt);
}
// Extra (twin) lattices: MultiLatticeSearch derives each rotation by relating the FFT's primary
// lattice[0] to its near-copies, so only apply it when the chosen cell IS that primary. If a
// widened alternative won (a superstructure/large cell), lattice[0] is a different (sublattice)
// metric and its rotations would misorient the chosen cell.
if (have_best && best_ci == 0 && indexer_result.lattice.size() > 1) {
auto ml_latt = MultiLatticeSearch(indexer_result.lattice);
for (auto &l : ml_latt) {
if (extra_lattices_.size() >= experiment.GetIndexingSettings().GetMaxExtraLattices())
break;
// Ignore lattices oriented by less than 3.0 degree
if (l.rotation_vector.Length() < 3.0 * PI / 180.0)
continue;
RotMatrix rot(l.rotation_vector.Length(), l.rotation_vector.Normalize());
XtalOptimizerData data_multi{
.geom = experiment_copy.GetDiffractionGeometry(),
.latt = indexed_lattice->Multiply(rot),
.crystal_system = search_result_.system,
.min_spots = experiment.GetIndexingSettings().GetViableCellMinSpots(),
.refine_beam_center = false,
.refine_detector_angles = false,
.refine_unit_cell = false,
.refine_rotation_axis = false,
.index_ice_rings = experiment.GetIndexingSettings().GetIndexIceRings(),
.max_iterations = real_time ? 0 : ROT_REFINE_ITERATIONS,
.axis = axis_
};
// Quick refinement: orientation only. Cell size/angles, beam center,
// detector angles and rotation axis are all kept from the first lattice.
// XtalOptimizer always refines orientation; everything else is frozen above.
XtalOptimizer(data_multi, v_);
extra_lattices_.push_back(data_multi.latt);
}
}
}
}
void RotationIndexer::ProcessImage(int64_t image, const std::vector<SpotToSave> &spots,
std::optional<float> angle_deg) {
std::unique_lock ul(m);
// For non-rotation just ignore the whole procedure
if (!axis_)
return;
// Guard: `image` is a slot in [0, image count); a bad index (e.g. a global number for a subset
// run) must not corrupt memory.
if (image < 0 || image >= static_cast<int64_t>(v_.size()))
return;
if (accumulated_spots >= max_spots)
return;
if (indexed_lattice)
return;
angle_deg_[image] = angle_deg;
v_[image].reserve(spots.size());
for (const auto &s: spots) {
if (index_ice_rings || !s.ice_ring)
v_[image].emplace_back(s);
}
// truncate spots, so we don't get above max_spots (total) and max_spots_per_image (for this image)
size_t max_spots_limit = std::min(max_spots_per_image, max_spots - accumulated_spots);
if (v_[image].size() > max_spots_limit) {
std::ranges::nth_element(v_[image], v_[image].begin() + max_spots_limit,
[](const SpotToSave &a, const SpotToSave &b) {
return a.intensity > b.intensity;
}
);
v_[image].resize(max_spots_limit);
}
accumulated_spots += v_[image].size();
}
std::vector<Coord> RotationIndexer::AccumulateReciprocal(const DiffractionGeometry &g,
const std::optional<GoniometerAxis> &ax) const {
std::vector<Coord> c;
c.reserve(accumulated_spots);
for (int i = 0; i < v_.size(); i++) {
const float a = angle_deg_[i].value_or(ax->GetAngle_deg(i) + ax->GetWedge_deg() / 2.0f);
const auto rot = ax->GetTransformationAngle(a);
for (const auto &s : v_[i])
c.emplace_back(rot * s.ReciprocalCoord(g));
}
return c;
}
std::optional<RotationIndexer::ConstrainedRefit>
RotationIndexer::RefineConstrained(const CrystalLattice &latt, gemmi::CrystalSystem system) const {
std::unique_lock ul(m);
if (!axis_ || accumulated_spots == 0 || latt.CalcVolume() <= 1.0)
return {};
const UnitCell uc = latt.GetUnitCell();
XtalOptimizerData d{
.geom = updated_geom_,
.latt = latt,
.crystal_system = system == gemmi::CrystalSystem::Trigonal ? gemmi::CrystalSystem::Hexagonal
: system,
.min_spots = experiment.GetIndexingSettings().GetViableCellMinSpots(),
.max_length_A = 1.2f * std::max({uc.a, uc.b, uc.c}),
// The indexers' bound, so a monoclinic beta outside [60,120] is refined and not clamped.
.min_angle_deg = 30.0f,
.max_angle_deg = 150.0f,
// The detector and the axis stay where the indexing left them - see the header.
.refine_beam_center = false,
.refine_detector_angles = false,
.refine_rotation_axis = false,
.index_ice_rings = index_ice_rings,
.max_iterations = real_time ? 0 : ROT_REFINE_ITERATIONS,
.axis = axis_
};
// One cloud for both scores: the geometry is held, so the lattice that comes out is scored on the
// same spots the one that went in is.
const auto cloud = AccumulateReciprocal(updated_geom_, axis_);
const float tol = experiment.GetIndexingSettings().GetTolerance();
const float before = IndexedFraction(latt, cloud, tol);
if (!XtalOptimizer(d, v_, 4))
return {};
return ConstrainedRefit{
.lattice = d.latt,
.indexed_fraction = IndexedFraction(d.latt, cloud, tol),
.indexed_fraction_before = before,
};
}
std::optional<RotationIndexerResult> RotationIndexer::GetLattice() const {
std::unique_lock ul(m);
if (!indexed_lattice)
return {};
return RotationIndexerResult{
.lattice = indexed_lattice.value(),
.extra_lattices = extra_lattices_,
.search_result = search_result_,
.geom = updated_geom_,
.axis = axis_,
.unconstrained = unconstrained_,
};
}
std::optional<std::string> RotationIndexer::GetIndexerError() const {
std::unique_lock ul(m);
return indexer_error_;
}
void RotationIndexer::ForceResult(const RotationIndexerResult &result) {
std::unique_lock ul(m);
indexed_lattice = result.lattice;
extra_lattices_ = result.extra_lattices;
search_result_ = result.search_result;
updated_geom_ = result.geom;
axis_ = result.axis;
unconstrained_ = result.unconstrained;
}
bool RotationIndexer::AccumulationFull() const {
std::unique_lock ul(m);
return accumulated_spots >= max_spots;
}
void RotationIndexer::ForceLattice(const CrystalLattice &lattice) {
indexed_lattice = lattice;
const gemmi::SpaceGroup &sg = experiment.GetSpaceGroupOrP1();
search_result_ = LatticeSearchResult{
.niggli_class = 0, // Since Niggli class was not searched for, we don't know which one
.conventional = lattice, // If lattice provided, it is for now primitive == conventional
.system = sg.crystal_system(),
.centering = sg.centring_type(),
};
}