WorksPhysics and Astronomy
From Autonomous Knowledge Generation to Generalizable Property Prediction: A Precision-First Pipeline for Electro-Optic Materials Discovery
Abstract
Electro-optic materials, whose refractive index changes linearly with an applied electric field (the Pockels effect), set the speed and energy of optical modulators and microwave-to-optical links for quantum computing. Designing better materials requires optimizing composition alongside temperature, strain, and bias field. However, general repositories lack electro-optic tensor data. We present a pipeline that builds such a database from literature following the rule: "no value is better than a wrong value." The pipeline has five stages: physics-aware acquisition, layout-aware extraction via LLM and deterministic table parsers, a three-tier verification gate, cross-paper consensus using median and median absolute deviation, and a noise-robust surrogate model audited by leave-one-material-out (LOMO) testing. Physics enters directly: notation is mapped, point-group symmetry-forbidden Pockels components are rejected, equivalent labels are merged, and values carry provenance tags. In a run over 49 publications, the gate processed 816 candidates into 414 accepted, 215 held for review, and 187 rejected. Against a hand-certified gold standard of 303 values from 16 papers, the run achieved a precision of 0.51 and recall of 0.48 for exact record matches. Across four database snapshots, held-out materials where the surrogate model beat a predict-the-mean baseline grew from 1 of 4 to 9 of 13 as the database expanded from 10 to 22 materials. While not a controlled comparison, this trend supports the premise that a precision-oriented, physics-checked database can power forward models beyond its training set and clarifies requirements for inverse material design.
Introduction
Why electro-optic materials matter
An electro-optic material changes its refractive index when an electric field is applied to it. In the linear (Pockels) effect, the change is proportional to the field and is described by a third-rank tensor [15, 87, 86],
where are the six independent components of the optical indicatrix in contracted (Voigt) notation, is the applied field along crystal axis , and are the Pockels coefficients, usually quoted in pm/V. Which of the 18 components may be non-zero is fixed by the crystal’s point group, and only crystals without a centre of inversion have any. A phase modulator of length and electrode gap needs a half-wave voltage
where is the optical wavelength, the refractive index, the effective coefficient for the chosen field and polarisation geometry, and the overlap of the optical and electrical fields. A larger means a lower drive voltage, a smaller device, and less energy per bit, which is why lithium niobate and, more recently, barium titanate on silicon have become the workhorses of integrated photonics [74, 75, 76, 31]. Figure 1(b) shows the effect in a modulator.
[x=1mm, y=1mm, font=\footnotesize,
every node/.append style={execute at begin node={\hyphenpenalty=10000\relax}},
mw/.style={DDNavy, line width=0.9pt, decorate,
decoration={snake, amplitude=1.6mm, segment length=7mm}},
opt/.style={DDPurple, line width=0.9pt, decorate,
decoration={snake, amplitude=1.1mm, segment length=2.6mm}},
lab/.style={font=\scriptsize, text=DDMeta, align=center},
head/.style={font=\footnotesize\bfseries, align=center}]
%% ---------------- panel (a) ----------------------------------------------
\node[anchor=west, font=\small\bfseries, text=DDNavy] at (0,64) {(a)};
\node[anchor=west, font=\small, text=DDInk] at (6,64)
{The electro-optic link of a networked quantum processor};
% cryostat
\draw[DDNavy!70, dashed, rounded corners=2mm, fill=DDNavy!3]
(0,8) rectangle (97,58);
\node[anchor=north west, font=\scriptsize\bfseries, text=DDNavy]
at (1.5,57) {Cryostat: millikelvin to a few kelvin};
% qubit chip
\draw[DDNavy, fill=white, rounded corners=1mm] (5,20) rectangle (31,44);
\foreach \x in {10,18,26} {
\draw[DDNavy, line width=0.7pt] (\x-2.2,32-2.2) -- (\x+2.2,32+2.2);
\draw[DDNavy, line width=0.7pt] (\x-2.2,32+2.2) -- (\x+2.2,32-2.2);
\draw[DDNavy, line width=0.7pt] (\x,26) -- (\x,29.8);
\draw[DDNavy, line width=0.7pt] (\x,34.2) -- (\x,38);
}
\node[head, text=DDNavy] at (18,48.5) {Qubits};
\node[lab] at (18,15) {superconducting\\circuits, $\sim$10\,mK};
% microwave photon
\draw[mw] (32,32) -- (51,32);
\draw[-{Stealth[length=2mm]}, DDNavy] (51,32) -- (54,32);
\node[lab, text=DDNavy] at (43,40) {microwave\\photon, $\sim$5\,GHz};
% transducer: microwave resonator electrodes around EO crystal + optical cavity
\fill[DDTeal!10] (55,24) rectangle (89,40);
\draw[DDTeal, line width=0.6pt] (55,24) rectangle (89,40);
\fill[DDGold!65] (57,40) rectangle (87,42.2);
\fill[DDGold!65] (57,21.8) rectangle (87,24);
\foreach \x in {61,67,...,85} {
\draw[-{Stealth[length=1.4mm]}, DDOrange!85, line width=0.5pt] (\x,39) -- (\x,34);
}
\node[font=\scriptsize, text=DDTeal!80!black] at (72,29.5)
{electro-optic crystal, $r_{ij}$};
\node[font=\scriptsize, text=DDOrange!85!black, anchor=west] at (88.5,37.5) {$E$};
\node[head, text=DDTeal!80!black] at (72,48.5) {Transducer};
\node[lab] at (68,15) {$\Delta(1/n^{2})_i = r_{ij}E_j$ mixes the\\microwave
field into the light};
% optical pump in, optical photon out
\draw[-{Stealth[length=2mm]}, DDRed!80, line width=0.8pt] (85,2) -- (85,21.5);
\node[lab, text=DDRed!80!black, anchor=west] at (86,4) {optical pump};
% optical photon out through fibre
\draw[opt] (90,32) -- (122,32);
\draw[-{Stealth[length=2mm]}, DDPurple] (122,32) -- (125,32);
\node[lab, text=DDPurple] at (110,40) {optical photon,\\$\sim$194\,THz (1550\,nm)};
\node[lab] at (110,22) {low-loss fibre,\\room temperature};
% network
\draw[DDPurple, fill=DDPurple!5, rounded corners=1mm] (126,20) rectangle (163,44);
\node[head, text=DDPurple] at (144.5,48.5) {Optical network};
\node[lab, text=DDInk] at (144.5,32) {other quantum\\processors,\\memories, detectors};
%% ---------------- panel (b) ----------------------------------------------
\begin{scope}[yshift=-66mm]
\node[anchor=west, font=\small\bfseries, text=DDNavy] at (0,55) {(b)};
\node[anchor=west, font=\small, text=DDInk] at (6,55)
{The Pockels effect: a field changes the index and shifts the phase};
% input light
\draw[DDRed!80, line width=0.9pt, domain=0:34, samples=120, smooth]
plot (\x, {34 + 5*sin(\x/6.5*360)});
\node[lab, text=DDRed!80!black] at (17,45) {input light};
\node[lab] at (17,24) {$E_0\cos(\omega t)$};
\draw[-{Stealth[length=2mm]}, DDRed!80, line width=0.8pt] (35,34) -- (44,34);
% crystal with electrodes
\fill[DDTeal!10] (45,26) rectangle (105,42);
\draw[DDTeal, line width=0.6pt] (45,26) rectangle (105,42);
\fill[DDGold!65] (45,42) rectangle (105,45);
\fill[DDGold!65] (45,23) rectangle (105,26);
\node[font=\scriptsize, text=DDGold!40!black] at (75,47.5) {electrode, $+V$};
\node[font=\scriptsize, text=DDGold!40!black] at (75,20.5) {electrode, ground};
\foreach \x in {50,58,...,98} {
\draw[-{Stealth[length=1.4mm]}, DDOrange!85, line width=0.5pt] (\x,41) -- (\x,36);
}
\draw[DDRed!70, line width=0.6pt, dashed] (45,31) -- (105,31);
\node[font=\scriptsize, text=DDTeal!80!black] at (75,28.3)
{$n \;\rightarrow\; n + \Delta n$};
\draw[{Stealth[length=1.6mm]}-{Stealth[length=1.6mm]}, DDMeta] (45,16) -- (105,16);
\node[lab, fill=white, inner sep=1pt] at (75,16) {length $L$};
\draw[{Stealth[length=1.4mm]}-{Stealth[length=1.4mm]}, DDMeta] (108,26) -- (108,42);
\node[lab, anchor=west] at (109,34) {$d$};
% output light, phase shifted
\draw[-{Stealth[length=2mm]}, DDRed!80, line width=0.8pt] (113,34) -- (122,34);
\draw[DDRed!45, line width=0.7pt, dashed, domain=124:163, samples=120, smooth]
plot (\x, {34 + 5*sin((\x-124)/6.5*360)});
\draw[DDPurple, line width=0.9pt, domain=124:163, samples=120, smooth]
plot (\x, {34 + 5*sin((\x-124)/6.5*360 - 70)});
\node[lab, text=DDPurple] at (143.5,45) {output light};
\node[lab] at (143.5,24) {$E_0\cos(\omega t + \Delta\varphi)$\\[-0.3mm]{\color{DDRed!60}dashed: without field}};
% relations
\node[draw=DDRule, fill=DDLight, rounded corners=1mm, inner sep=2.5mm,
font=\footnotesize, align=center] at (81.5,4)
{$\Delta n = -\tfrac12\, n^{3} r_{\mathrm{eff}} E$, \quad
$\Delta\varphi = \dfrac{2\pi}{\lambda}\,\Delta n\, L$, \quad
$E = V/d$ \qquad$\Longrightarrow$\qquad
larger $n^{3}r_{\mathrm{eff}}$: lower $V_\pi$, less energy per bit, shorter devices};
\end{scope}Figure 1. Why electro-optic materials matter. (a) If qubits are the brain of a quantum computer, electro-optic materials are its nervous system. Superconducting qubits work at millikelvin temperatures and speak in microwave photons, whereas low-loss transmission over distance uses optical photons. A cavity electro-optic transducer couples the two through the Pockels effect of a crystal held at cryogenic temperature [72, 73, 79, 78]; its efficiency and added noise depend on the Pockels coefficient the crystal keeps at a few kelvin. (b) In a phase modulator the applied field changes the refractive index by ((1)) and shifts the phase of the light by . A larger lowers the half-wave voltage ((2)) and the switching energy, which scales as for electrode capacitance .
Electro-optic, optical, and nonlinear-optical materials
These three terms are often used loosely, and the distinction matters for what the database stores. Optical materials is the broad class of materials used to transmit, guide, or shape light, characterised by refractive index, absorption, and dispersion. Nonlinear-optical materials respond to the optical field itself in a way that is not linear in that field; second-order () processes such as second-harmonic generation and parametric down-conversion mix optical frequencies with each other. Electro-optic materials are those whose optical properties are controlled by an external, low-frequency or static electric field. The Pockels effect is the limit of the same response in which one of the two fields is at, or near, zero frequency, and the quadratic (Kerr) effect is its third-order counterpart [86]. A crystal can be an excellent nonlinear-optical material and still be a poor modulator, so the database records Pockels and effective electro-optic coefficients as their own property family and treats nonlinear-optical susceptibilities as out of scope.
The nervous system of a quantum computer
If quantum materials are the brain of a quantum computer, the media in which qubits are made and controlled [84], electro-optic materials are its nervous system. Superconducting qubits operate at millikelvin temperatures and are controlled and read out by microwave signals, whereas low-loss communication over distance is optical. Linking processors into a network, or even moving their signals out of a dilution refrigerator without a bundle of heat-carrying coaxial cables, needs a coherent converter between microwave and optical photons [72, 73] (Figure 1a). Cavity electro-optic transducers do this through the Pockels effect [79, 78], and cryogenic electro-optic links already read out superconducting circuits optically [80]. In each case the conversion efficiency, the added noise, and the heat load are set by the Pockels coefficient that the material retains at a few kelvin, and that coefficient is not the room-temperature value printed in handbooks. In barium titanate it drops by roughly a factor of two on cooling but stays far above lithium niobate [77]; in strontium titanate, a quantum paraelectric, thin films engineered for low temperature reach very high values [60]; and epitaxial strain changes it again [56, 61].
The design problem: a material and its knobs
The Pockels response is therefore not a single number per compound. It depends on the composition and phase, on the crystal or film form and the substrate, and on external knobs: temperature, mechanical strain or clamping, a bias field that can induce a large linear response from a quadratic one (as in potassium tantalate niobate [46]), and the optical wavelength and modulation frequency. Designing an electro-optic material means searching jointly over the material configuration \vect{c} and the operating conditions \vect{k},
for example the configuration with the largest Pockels coefficient at 4 K under a strain the substrate can supply, where is temperature, strain, the bias field, the wavelength, and the modulation frequency. Figure 2 sketches this joint space.
[x=1mm, y=1mm, font=\footnotesize,
every node/.append style={execute at begin node={\hyphenpenalty=10000\relax}},
lab/.style={font=\scriptsize, text=DDMeta, align=center},
head/.style={font=\footnotesize\bfseries, align=center},
panel/.style={draw=#1, fill=#1!4, rounded corners=1.5mm, line width=0.6pt},
atA/.style={circle, fill=DDNavy!70, inner sep=0pt, minimum size=2.6mm},
atB/.style={circle, fill=DDRed!75, inner sep=0pt, minimum size=2.4mm},
atO/.style={circle, draw=DDGray, fill=white, inner sep=0pt, minimum size=1.6mm}]
%% ---------- material configuration -------------------------------------
\draw[panel=DDNavy] (0,0) rectangle (66,62);
\node[head, text=DDNavy] at (33,57.5) {Material configuration $\vect{c}$};
% perovskite ABO3 cell, [001] projection: BO6 octahedron (diamond) + A corners
\begin{scope}[shift={(17,36)}]
\fill[DDRed!8] (0,8) -- (8,0) -- (0,-8) -- (-8,0) -- cycle;
\draw[DDGray, line width=0.4pt] (0,8) -- (8,0) -- (0,-8) -- (-8,0) -- cycle;
\draw[DDGray!70, line width=0.3pt] (-8,0) -- (8,0) (0,-8) -- (0,8);
\draw[DDNavy!50, line width=0.4pt] (-11,-11) rectangle (11,11);
\foreach \p in {(-11,-11),(11,-11),(11,11),(-11,11)} \node[atA] at \p {};
\node[atB] at (0,0) {};
\foreach \p in {(0,8),(8,0),(0,-8),(-8,0)} \node[atO] at \p {};
\node[font=\scriptsize, text=DDNavy] at (-11,-14.5) {A};
\node[font=\scriptsize, text=DDRed!80!black] at (3.2,2.6) {B};
\node[font=\scriptsize, text=DDMeta] at (6.2,-3.6) {O};
\end{scope}
\node[lab] at (17,19.5) {perovskite ABO$_3$};
\node[anchor=north west, align=left, font=\scriptsize, text=DDInk,
text width=34mm] at (31,51)
{\textbf{Simple:} ABO$_3$\\[1.2mm]
\textbf{Solid solutions:}\\(A,A$'$)(B,B$'$)O$_3$\\[1.2mm]
\textbf{Superlattices:}\\(ABO$_3$)$_m$/(A$'$B$'$O$_3$)$_n$\\[1.2mm]
\textbf{Layered:} Ruddlesden--Popper, Dion--Jacobson, Aurivillius};
\node[anchor=south, align=center, font=\scriptsize, text=DDInk,
text width=62mm] at (33,2)
{also niobates, tantalates, phosphates, nitrides, and zinc-blende
semiconductors; phase, orientation, domain state, film or bulk};
%% ---------- times ---------------------------------------------------------
\node[font=\LARGE, text=DDMeta] at (70.5,31) {$\times$};
%% ---------- operating knobs ----------------------------------------------
\draw[panel=DDOrange] (75,0) rectangle (125,62);
\node[head, text=DDOrange!85!black] at (100,57.5) {Operating knobs $\vect{k}$};
% temperature
\begin{scope}[shift={(83,44)}]
\draw[DDOrange!85, line width=0.6pt, rounded corners=0.8mm] (-1.2,-3) rectangle (1.2,6);
\fill[DDNavy!70] (0,-4.2) circle (2.1mm);
\fill[DDNavy!70] (-0.6,-3) rectangle (0.6,0.5);
\end{scope}
\node[anchor=west, font=\scriptsize, text=DDInk, align=left] at (88,45)
{temperature $T$:\\millikelvin to room temperature};
% strain
\begin{scope}[shift={(83,29)}]
\draw[DDTeal, fill=DDTeal!10] (-3,-3) rectangle (3,3);
\draw[-{Stealth[length=1.2mm]}, DDTeal] (3.6,0) -- (6.2,0);
\draw[-{Stealth[length=1.2mm]}, DDTeal] (-3.6,0) -- (-6.2,0);
\draw[-{Stealth[length=1.2mm]}, DDTeal] (0,3.6) -- (0,6.2);
\draw[-{Stealth[length=1.2mm]}, DDTeal] (0,-3.6) -- (0,-6.2);
\end{scope}
\node[anchor=west, font=\scriptsize, text=DDInk, align=left] at (90,29)
{strain $\varepsilon$: biaxial epitaxial\\(substrate) or uniaxial};
% electric field
\begin{scope}[shift={(83,14)}]
\fill[DDGold!65] (-4,3) rectangle (4,4.2);
\fill[DDGold!65] (-4,-4.2) rectangle (4,-3);
\foreach \x in {-2.5,0,2.5}
\draw[-{Stealth[length=1.1mm]}, DDOrange!85] (\x,2.6) -- (\x,-2.6);
\end{scope}
\node[anchor=west, font=\scriptsize, text=DDInk, align=left] at (90,14)
{bias field $E_{\mathrm{bias}}$ (can induce\\a linear response)};
\node[anchor=south, font=\scriptsize, text=DDInk] at (100,2)
{wavelength $\lambda$, modulation frequency $f$};
%% ---------- arrow to response -------------------------------------------
\draw[-{Stealth[length=2.4mm]}, DDNavy, line width=1pt] (126.5,31) -- (133,31);
\draw[panel=DDTeal] (134,0) rectangle (165,62);
\node[head, text=DDTeal!80!black] at (149.5,57.5) {Response};
\node[font=\normalsize, text=DDTeal!80!black] at (149.5,46)
{$n^{3}r_{\mathrm{eff}}(\vect{c},\vect{k})$};
\node[lab, text width=28mm] at (149.5,35)
{target: maximum at the operating point, e.g.\ 4\,K and a given strain};
\draw[DDRule] (137,25) -- (162,25);
\node[lab, text width=29mm, text=DDRed!80!black] at (149.5,13.5)
{The Materials Project and OQMD hold no $r_{ij}$ and no measurement conditions; the data exist only in papers.};Figure 2. The electro-optic design problem of (3). The response depends jointly on the material configuration, which for perovskite oxides alone spans solid solutions, superlattices, and layered phases, and on external knobs that can each change the Pockels coefficient substantially: temperature [77, 60], strain [56, 61], and bias field [46]. The joint space is far too large to search by experiment alone, and a model that could search it needs condition-resolved Pockels data that no general materials database contains. Building those data is the subject of this paper.
Why there is little AI for electro-optics
Machine learning has accelerated discovery where large, consistent property databases exist. The Materials Project [1], the Open Quantum Materials Database (OQMD) [2], AFLOW [3], and the Inorganic Crystal Structure Database (ICSD) [4] standardise structures, formation energies, and electronic descriptors, and the Materials Project also holds computed elastic, piezoelectric, and dielectric tensors [81, 82, 83]. None of them holds electro-optic tensors, and none records the temperature, strain, field, and wavelength at which a functional property was measured. Those data exist only in papers, reported under incompatible conventions (, , effective coefficients, clamped and unclamped values) and printed alongside dielectric, elastic, and piezoelectric constants that are easily confused with them. Without a database, (3) cannot be posed to a model at all. Building that database, and showing that it supports prediction beyond the materials it contains, is the purpose of this paper.
The data bottleneck
The knowledge that governs functional behaviour remains largely in unstructured publications: the synthesis route, crystallographic orientation, phase, measurement method, and the conditions of measurement [5, 6, 7]. A property value has scientific meaning only when it is bound to its material, its tensor component, and its conditions. Automated literature mining has become practical. Rule-based toolkits such as ChemDataExtractor [8] have produced databases of magnetic phase-transition temperatures [9] and battery materials [10], and text mining has recovered synthesis routes [11]. Large language models now extract structured records from scientific text with accuracy approaching that of expert curators [12, 13]. The same models can also attach a correct number to the wrong label, material, or condition, or produce a value that is not in the source [14]. For a tensor property, where one wrong component can corrupt every model trained on it, that failure matters more than coverage.
In earlier work we introduced an autonomous scientific knowledge generation framework that turns literature into an AI-ready knowledge base through ontology-guided acquisition, hybrid extraction, and semantic harmonisation [22]. That work set out the architecture on a small corpus and left two questions open. First, how accurate are the values the system emits, and can a verification step stop the physically impossible ones? Second, does the resulting database support prediction for materials held out of training, which is the property any generative design step depends on? This paper answers both for inorganic electro-optic materials. A pipeline that extracts their tensors correctly has to reason about symmetry, notation, and measurement context, which are the same capabilities needed for batteries, catalysis, and quantum materials.
The contributions of this paper are:
A five-stage pipeline (physics-aware acquisition, layout-aware extraction, a three-tier verification gate, cross-paper consensus, and noise-robust modelling) in which each stage can refuse to pass a value forward, with a flowchart and an algorithm for each stage (sec:framework,sec:methods).
Physics embedded in verification: reversible symbol recovery that maps to , rejection of Pockels components forbidden in cubic zinc-blende crystals, point-group merging of symmetry-equivalent labels, and provenance tags that separate measured, computed, and device-effective values (§ 3.3).
A hand-certified gold standard drawn from 16 papers (§ 3.6) and the measured accuracy of the pipeline against it, reported with the gate’s own tallies, so that precision is a number rather than a design intention (§ 4.2).
Evidence that leave-one-material-out prediction improves as the number of distinct materials grows, from 1 of 4 to 9 of 13 held-out materials across four database snapshots (§ 4.4), and a concrete path from this forward model to inverse design over materials and operating conditions (§ 5).
A companion paper [23] analyses the errors that survive this pipeline, with a taxonomy, a layered detection framework, and an audit of accepted values against reference tensors. The two papers share one production run and use the same accuracy figures.
Framework
The pipeline
The pipeline converts a set of publications into a set of canonical property records through five stages (fig:pipeline). Stage 0 finds and ranks candidate papers before any are read; Stages 1 to 4 extract, verify, reconcile, and model. Each of Stages 1 to 4 may withhold a value rather than emit an uncertain one. § 3 describes each stage with its own flowchart.
Records, scope, and the acceptance rule
The unit the pipeline handles is a record, not a bare number. A record is the tuple
where is the material (for example LiNbO), the canonical tensor label (for example ), the numerical value, its unit, \vect{c} the measurement conditions (temperature, strain or mechanical boundary, wavelength, frequency, form, substrate), the provenance (experiment, first-principles calculation, fitted set, derived, or unspecified), and the source (paper and page). The database targets inorganic crystalline electro-optic, ferroelectric, and related materials; organic and hybrid systems, whose conventions and symmetry warrant a separate schema, are routed out at detection. Within scope, one acceptance rule is applied.
where holds if the value is physically admissible for the material’s symmetry and property family, if the record carries the context that makes it interpretable and is the paper’s own result, and if its symbol has been resolved to a canonical label. A record that fails any clause goes to review or is rejected, never into the database. Recall is recovered later, by consensus and by a modelling stage that tolerates residual noise, rather than by lowering the bar for acceptance.
Methods
Stage 0: physics-aware literature acquisition
Stage 0 (Figure 3) issues bibliographic queries to Crossref, arXiv, and PubMed so that coverage does not depend on a single index. Each query pairs a target material or an electro-optic term with a context term. Returned records are de-duplicated by digital object identifier (DOI), or by normalised title when no DOI exists, against a persistent SQLite catalogue. A relevance ranker then scores each record from its title, abstract, and keywords.
[node distance=6mm and 5mm]
\node[stnode, ddevidence] (q) {Query plan\\{\scriptsize materials $\times$ EO terms}};
\node[stnode, ddmethod, right=of q] (r) {Retrieve\\{\scriptsize Crossref, arXiv, PubMed}};
\node[stnode, ddmethod, right=of r] (d) {De-duplicate\\{\scriptsize DOI or normalised title}};
\node[stnode, ddmethod, right=of d] (s) {Score relevance\\{\scriptsize $R(p)$, \cref{eq:relevance}}};
\node[stdec, below=9mm of s] (t) {$R(p)\ge\tau$?};
\node[stnode, ddmethod, left=of t] (dl) {Download\\{\scriptsize open-access PDF}};
\node[stnode, ddmethod, left=of dl] (v) {Validate file\\{\scriptsize PDF header, SHA-256}};
\node[stnode, ddresult, left=of v] (cat) {Catalogue\\{\scriptsize metadata, score, embedding}};
\node[stnode, ddcontrol, below=7mm of t] (skip) {Skipped, logged\\{\scriptsize score kept}};
\draw[ddflow] (q) -- (r);
\draw[ddflow] (r) -- (d);
\draw[ddflow] (d) -- (s);
\draw[ddflow] (s) -- (t);
\draw[ddgate] (t) -- node[stlab, above]{yes} (dl);
\draw[ddflow] (dl) -- (v);
\draw[ddflow] (v) -- (cat);
\draw[ddreject] (t) -- node[stlab, right]{no} (skip);Figure 3. Stage 0, acquisition. Queries go to three bibliographic services; records are de-duplicated against the catalogue, scored by (5), and downloaded only above the threshold . A file is catalogued only if it is a valid PDF, and its SHA-256 hash detects byte-identical duplicates.
The score rewards direct electro-optic evidence and penalises vocabulary that shares surface words but not the target physics. For a paper , let be the number of matches in its title, abstract, and keywords to pattern group : direct electro-optic terms (“electro-optic”, “Pockels”, explicit coefficients such as ), target crystals, ferroelectric physics, thin-film growth, quantum-device context (cryogenic, transduction, qubit), and computational electro-optics. Then
where the group weights are (direct electro-optic), 4.0 (target material), 2.0 (ferroelectric), 1.7 (thin film), 2.5 (quantum device), and 1.2 (computational), each count is capped at (5 for ferroelectric) so that one repeated word cannot dominate, and the bonuses reward the co-occurrences that identify a target paper: electro-optic term with target material (), material with ferroelectric physics (), material with thin film (), electro-optic term with quantum device (), and electro-optic term with computation (). counts generic nonlinear-optics terms (second-harmonic generation, parametric oscillation and down-conversion, squeezing) with penalty , scaled by when the paper also has a direct electro-optic term and a target material, 0.6 when it has one of the two, and 1 otherwise; counts out-of-scope terms with penalty . Here is the raw score and equals 1 when its condition holds and 0 otherwise. Two gates override the sum: a record with no direct electro-optic, material, ferroelectric, or quantum-device signal scores zero, and a record whose only signal is generic nonlinear optics is capped at a raw score of 2. The divisor 30 is the raw score of a typical strong electro-optic materials paper (18 to 35 in practice), so reads as a fraction of that. Records with are downloaded from open-access sources; the ranked list shown to the user uses 0.20. Each stored paper keeps its metadata, score, keywords, and a text embedding for semantic search.
Stage 1: layout-aware extraction
Each publication is parsed page by page into text that preserves reading order across two-column layouts, with tables recovered as structured rows and page numbers kept for provenance (Figure 4). Text normalisation repairs ligatures and character encodings without splitting words or formulas.
[node distance=6mm and 5mm]
\node[stnode, ddevidence] (pdf) {PDF\\{\scriptsize from Stage 0}};
\node[stnode, ddmethod, right=of pdf] (parse) {Parse layout\\{\scriptsize reading order, tables, pages}};
\node[stnode, ddmethod, right=of parse] (det) {Detect material\\{\scriptsize \cref{alg:detect}}};
\node[stdec, right=of det] (conf) {In scope and confident?};
\node[stnode, ddcontrol, below=8mm of conf] (held) {Paper held\\{\scriptsize manual assignment}};
\node[stnode, ddmethod, below=8mm of det] (ctx) {Paper context\\{\scriptsize form, substrate, $T$, $f$}};
\node[stnode, ddmethod, below left=8mm and -4mm of ctx] (tab) {Table parsers\\{\scriptsize Pockels matrix, constant tables}};
\node[stnode, ddmethod, below right=8mm and -4mm of ctx] (llm) {LLM per page\\{\scriptsize GPT-4o, JSON schema}};
\node[stnode, ddresult, below=30mm of ctx] (merge) {Merge\\{\scriptsize deterministic row preferred}};
\node[stnode, ddresult, right=of merge] (cand) {Candidate records\\{\scriptsize to Stage 2}};
\draw[ddflow] (pdf) -- (parse);
\draw[ddflow] (parse) -- (det);
\draw[ddflow] (det) -- (conf);
\draw[ddreject] (conf) -- node[stlab, right]{no} (held);
\draw[ddgate] (conf.south west) -- node[stlab, above left]{yes} (ctx.north east);
\draw[ddflow] (ctx) -- (tab);
\draw[ddflow] (ctx) -- (llm);
\draw[ddflow] (tab) -- (merge);
\draw[ddflow] (llm) -- (merge);
\draw[ddflow] (merge) -- (cand);Figure 4. Stage 1, extraction. The subject material is identified first, because a wrong material label corrupts every value a paper contributes. A paper whose material cannot be identified with confidence is held and none of its values are extracted. Deterministic parsers and the LLM then read the paper in parallel, and their outputs are merged.
Subject-material detection
Before any value is extracted, the paper’s subject material is identified from a curated vocabulary of 47 inorganic materials (alg:detect). Each material is listed with its surface forms : the ways it is written in papers, such as formulas (LiNbO), names (lithium niobate), and abbreviations (LN). Detection is conservative, because a mislabelled material silently corrupts every value a paper contributes. Let be the full text of the paper, its early text (the first 2500 characters of page 1, which contain the title, authors, and abstract), and the title zone (the first 600 characters of ). With the number of whole-word matches of form in text , the score of material is
Because , a mention in the title counts times, a mention in the abstract 4 times, and a mention in the body once. The weights encode where a paper names its own subject. A paper on a new crystal routinely compares it with lithium niobate or barium titanate dozens of times in the body, but names its own material in the title, so makes one title mention outweigh about forty body mentions of a reference crystal. The abstract weight gives early prominence extra credit without letting a single comparison in the abstract (“larger than LiNbO”) decide the result. The set holds 26 ambiguous acronyms (such as KDP, BTO, STO, LN, PZT, KTN, and SiC) that also have unrelated meanings. They earn at most 0.3, which breaks ties between materials that already have a strong score but cannot create a winner on its own. The values of , , and 0.3 were set by inspecting detection errors on the corpus, not by optimisation.
Two guards protect the result. A gated abbreviation credits an abbreviation that a paper coins for its own crystal (for example “KTN” for KTaNbO) only when one of the confirming formulas also occurs in , so a coined abbreviation cannot hijack a paper about the parent compound. The confidence guard requires the winner to occur in the early text . A winner that never appears there is almost always a cited comparison in a paper whose true subject is outside the vocabulary (an organic or halide-perovskite paper, for example). Such a paper is held: it is placed on a list for manual assignment, and none of its values are extracted until a person assigns the material. A paper in which no material dominates ( for the runner-up ) is tagged as a comparison paper but is still processed, because the confidence guard has already established that its winner is named early.
Word-boundary anchoring stops short acronyms (“STO”, “LN”, “SiC”) from matching inside other words. A separate detector records paper-level context (form, substrate, temperature, frequency), which is later attached to each record.
Language-model extraction
Values are read from each page by a general-purpose LLM (GPT-4o, OpenAI) at temperature zero in JSON (JavaScript Object Notation) mode, under a schema that names the subject material, the allowed property families and tensor labels, and the condition fields to fill. In the production run the cost was under ten US cents per paper.
Deterministic table parsers
Two structures that free-text reading handles poorly are read by deterministic parsers. The first is the bracketed Pockels matrix, in which a full tensor is printed as a array of bare numbers whose indices are implied by position (as in [56]); the parser assigns each entry its label from its row and column. The second is a constant table of electro-optic, elastic, or refractive values with explicit subscripts. When a deterministic row and an LLM row report the same tensor and value, or the same tensor under the same condition, the deterministic row is kept. In the production run, 62 of the 414 accepted values came from these parsers and 352 from the LLM.
Stage 2: the three-tier verification gate
Every candidate record is assigned to one of three tiers (accepted, review, or rejected) by the conjunction of eq:accept (fig:stage2,alg:gate). The gate is positive: a record is accepted only if it matches a recognised label pattern for its property family, carries a unit consistent with that family, and passes the physics and symbol clauses below. A record that is not verified is kept with a typed reason, so that recall can be recovered by a later decision.
[node distance=5mm and 9mm]
\node[stwide, ddevidence] (in) {Candidate record $\rho$ from Stage 1};
\node[stwide, ddmethod, below=of in] (sym) {$S$: recover symbol\\{\scriptsize $\gamma\to r$, superscripts, merge by point group}};
\node[stdec, text width=36mm, below=of sym] (lab) {Label recognised; family, prefix, unit consistent?};
\node[stdec, text width=36mm, below=of lab] (phys) {$P$: in physical range and allowed by symmetry?};
\node[stdec, text width=36mm, below=of phys] (ctx) {$C$: paper's own result, verbatim, keyed, plausible?};
\node[stwide, ddresult, below=of ctx] (acc) {\textbf{Accepted}\\{\scriptsize provenance and condition completeness tagged}};
\node[ddnode, ddcontrol, font=\footnotesize, text width=58mm, right=of lab] (r1)
{\textbf{Rejected}: corrupted or index-only label, equation number\\
\textbf{Review}: missing indices, unit or family mismatch};
\node[ddnode, ddcontrol, font=\footnotesize, text width=58mm, right=of phys] (r2)
{\textbf{Rejected}: out of range, forbidden component (\cref{eq:zb}), out-of-scope family, OCR-corrupted};
\node[ddnode, ddcontrol, font=\footnotesize, text width=58mm, right=of ctx] (r3)
{\textbf{Review}: cited or comparison value, series value without key, not verbatim, implausible magnitude};
\draw[ddflow] (in) -- (sym);
\draw[ddflow] (sym) -- (lab);
\draw[ddgate] (lab) -- node[stlab, right]{yes} (phys);
\draw[ddgate] (phys) -- node[stlab, right]{yes} (ctx);
\draw[ddgate] (ctx) -- node[stlab, right]{yes} (acc);
\draw[ddreject] (lab) -- node[stlab, above]{no} (r1);
\draw[ddreject] (phys) -- node[stlab, above]{no} (r2);
\draw[ddreject] (ctx) -- node[stlab, above]{no} (r3);Figure 5. Stage 2, the verification gate. A record passes the symbol clause , a label and unit check, the physics clause , and the context clause in order, and is accepted only if it passes all of them. A record that is impossible (a corrupted label, a symmetry-forbidden component, an out-of-range value) is rejected; one that is merely unverified is held for review. Both keep their reasons.
Symbol recovery ()
Notation is normalised before it is judged. Underscores and run-together boundary superscripts are removed reversibly (, written as ), without contracting indices. The classical symbol for the linear electro-optic coefficient is mapped to , and combined coefficients to an effective coefficient . This rule recovered a five-component tensor from a hydrothermal KTiOPO (KTP) paper [48] that reported , , , , and ; the recovered \gamma_{33}=\SI{36.3}{pm/V} matched a value the same run had already accepted as r_{33}=\SI{36.3}{pm/V}.
Symmetry admissibility ()
Crystal symmetry fixes which components of in (1) may be non-zero [15]. For cubic zinc-blende crystals (point group ), namely GaAs, InP, GaP, ZnTe, ZnSe, CdTe and their alloys, the tensor has a single independent component:
Any other reported for such a material is rejected as an extraction error. Out-of-scope families (phonon frequencies, band gaps, optical loss) are rejected by the same guard (lst:gate). Each property family also has a physical range outside which a value is rejected: 0.05 pm/V–6000 pm/V for electro-optic coefficients, 0.05 to 3000 pC/N for piezoelectric strain coefficients, 1 to 2000 GPa for elastic constants, 1 to 200 000 for relative permittivity, and 1 to 6 for refractive index. The ranges are deliberately wide; they catch unit and decimal errors, not unusual materials.
ZB43M = {"GaAs","InP","GaP","ZnTe","ZnSe","CdTe","ZnxCd1-xTe","Cd1-xMgxTe"}
ZB_ALLOWED = {"r41", "r52", "r63", "r_eff"} # zinc-blende (-43m)
OUT_OF_SCOPE = {"phonon_frequency", "band_gap", "optical_loss", "raman"}
def admissible(material, family, label):
if family in OUT_OF_SCOPE: # not a database property
return False
if material in ZB43M and family == "electro_optic" \
and label not in ZB_ALLOWED: # symmetry-forbidden component
return False
return True
Symmetry also governs how a correctly read value is labelled. Two labels that the point group forces to be equal are merged into one canonical label, a step we call folding. Writing for the folding map of point group ,
applied to the trigonal crystals of the corpus (LiNbO, LiTaO, BiFeO, KNbBO), the tetragonal crystals (BaTiO, PIN–PMN–PT, SrTiO), and zinc-blende. For crystals and are kept as separate labels to match the reference tables the gold standard follows; for orthorhombic KNbO no components are equivalent and none are folded. On the electro-optic part of the gold standard, folding raised recall from 0.593 to 0.658 while precision was essentially unchanged (0.478 to 0.475). Folding relabels; it does not reject. A component that is forbidden in a non-cubic class is not yet rejected (§ 4.6).
Unit consistency
A value is admitted for a property family only if the unit reported with it is dimensionally consistent with that family. The rule was added after a permittivity of several thousand, read from a dense multi-property table whose column header sat several rows away, was assigned to an elastic-constant slot. An elastic value whose source text carries no elastic unit is now held.
Provenance and context ()
Each accepted record is tagged as experiment, first-principles calculation, fitted self-consistent set, derived relation, or unspecified. Device-level coefficients are labelled and marked , apart from intrinsic tensor components. Values the source cites from earlier work are held, so that a well-known constant is not counted again in every paper that quotes it. A value from a series (for example, one of several temperatures) without the key that identifies its member is held, as is a value that does not occur verbatim in the source text. Conditions are scored as complete, partial, or missing, and a bulk single crystal is marked as the zero-strain reference so that “unstrained” is distinguishable from “strain not reported”.
Stage 3: cross-paper consensus
Several papers often report the same quantity, for example of LiNbO at room temperature. Consensus here means the agreement among the independent reports of one quantity: a value that agrees with the other papers is more likely to be a correct reading than one that stands alone or disagrees. Stage 3 (fig:stage3,alg:consensus) measures that agreement and turns it into a confidence weight for each record.
[node distance=6mm and 5mm]
\node[stnode, ddevidence] (in) {Accepted records\\{\scriptsize from Stage 2}};
\node[stnode, ddmethod, right=of in] (grp) {Group by key $\kappa(\rho)$\\{\scriptsize \cref{eq:groupkey}}};
\node[stdec, right=of grp] (n3) {Group size $n\ge3$?};
\node[stnode, ddmethod, right=of n3] (mad) {Median, MAD, $z_i$\\{\scriptsize \cref{eq:mad}}};
\node[stdec, below=9mm of mad] (out) {$|z_i|>3.5$ and $\delta_i>0.15$?};
\node[stnode, ddresult, left=of out] (inl) {Inlier\\{\scriptsize $w=\min(1,0.5+0.1n)\,a$}};
\node[stnode, ddresult, below=9mm of n3] (single) {Single source\\{\scriptsize $w=0.5$}};
\node[stnode, ddcontrol, below=7mm of out] (rev) {Consensus review\\{\scriptsize $w\le0.2$, kept}};
\draw[ddflow] (in) -- (grp);
\draw[ddflow] (grp) -- (n3);
\draw[ddgate] (n3) -- node[stlab, above]{yes} (mad);
\draw[ddflow] (n3) -- node[stlab, right]{no} (single);
\draw[ddflow] (mad) -- (out);
\draw[ddgate] (out) -- node[stlab, above]{no} (inl);
\draw[ddreject] (out) -- node[stlab, right]{yes} (rev);Figure 6. Stage 3, consensus. Accepted records that describe the same quantity are grouped; in groups of three or more, a robust score identifies values that disagree with the rest. Every record leaves this stage with a confidence weight; none is deleted.
Records are grouped by the key
where is the property family and a coarse temperature regime: cryogenic below 150 K, room temperature up to 400 K, and 100 K bins above. The mechanical boundary condition (clamped or unclamped) can be added to the key; this split is off by default because many papers do not state it. Within a group of reports, agreement is assessed with the modified -score of Iglewicz and Hoaglin [16, 17], built on the median and the MAD, which are insensitive to the outliers they must flag:
where is the group median and the factor 0.6745 makes comparable to a standard normal score. A record is an outlier when and its relative deviation exceeds 0.15; when the MAD is zero (most reports identical), the condition alone applies. The second condition stops a group of nearly identical values from flagging a report that differs only in the last digit. The confidence weight is
where is the fraction of inliers in the group, so a fully agreeing group of three gives and of five or more gives .
The low weights () are chosen to keep a disputed record in the database without letting it steer the model. A record that disagrees with its group may be a transcription error, or it may be a correct but unusual measurement, for example a different domain state or a different wavelength that the group key does not separate. Frequency alone cannot tell the two apart, so deleting outliers would remove exactly the measurements that might matter most. With , a disputed record contributes at most a fifth of the pull of a verified one in the weighted loss of Stage 4, and less than half that of a single-source record (), whose only weakness is that nobody else has measured it. The ordering, not the exact values, is what matters: agreed single source disputed impossible. Outliers and unit conflicts are also placed on a consensus review list.
Stage 4: noise-robust modelling and the generalization test
The consensus database feeds a forward surrogate model that predicts for electro-optic records (Figure 7). The inputs \vect{x} are of three kinds: composition features (fraction-weighted mean and range of atomic number, atomic mass, electronegativity, and atomic radius, plus the number of elements), so that the model can score a formula absent from training; condition features (temperature in kelvin, strain in percent, wavelength, clamped or unclamped boundary, film or bulk, effective or intrinsic, computed or measured); and the canonical tensor label.
[node distance=6mm and 5mm]
\node[stnode, ddevidence] (db) {Weighted records\\{\scriptsize from Stage 3}};
\node[stnode, ddmethod, right=of db] (feat) {Features $\vect{x}$\\{\scriptsize composition, conditions, label}};
\node[stnode, ddmethod, right=of feat] (wt) {Row weights $\omega_i$\\{\scriptsize \cref{eq:weight}}};
\node[stnode, ddmethod, right=of wt] (fit) {Fit $f_\theta$\\{\scriptsize gradient boosting, Huber loss}};
\node[stnode, ddmethod, below=9mm of fit] (oof) {5-fold audit\\{\scriptsize residual $z_i$, \cref{eq:oof}}};
\node[stnode, ddmethod, below=9mm of wt] (lomo) {LOMO loop\\{\scriptsize hold out each material}};
\node[stnode, ddresult, left=of lomo] (dom) {Verdicts, domain $\Set{D}_\tau$\\{\scriptsize \cref{eq:lomo}}};
\node[stnode, ddcontrol, below=7mm of oof] (rev) {Review queue\\{\scriptsize $|z_i|>2.5$}};
\draw[ddflow] (db) -- (feat);
\draw[ddflow] (feat) -- (wt);
\draw[ddflow] (wt) -- (fit);
\draw[ddflow] (fit) -- (oof);
\draw[ddflow] (fit) -- (lomo);
\draw[ddflow] (lomo) -- (dom);
\draw[ddreturn] (oof) -- (rev);Figure 7. Stage 4, modelling. A robust regressor is trained on weighted records. An out-of-fold audit returns records the model finds incompatible with their neighbours to review, and the leave-one-material-out (LOMO) loop produces a per-material verdict that defines where the model may be trusted.
Each row is weighted by its consensus weight and by how complete its conditions are,
where is the consensus weight of (10). The regressor is gradient boosting [18] (300 trees, depth 3, learning rate 0.05, subsample 0.8) trained on the weighted Huber loss [19]
with set at each boosting step to the 90th percentile of the absolute residuals. The loss is quadratic for ordinary errors and linear for large ones, so a residual extraction error pulls the fit far less than it would under squared error.
The model then audits the labels it was trained on. A five-fold out-of-fold prediction is made for every record, from a model that did not see it, and the residuals are standardised robustly,
with the median residual. Records with are listed as incompatible with their chemical neighbours [20]. The list is a review queue, not a deletion rule. An ensemble of eight bootstrap random forests supplies a second signal, the spread \sigma_\theta(\vect{x}) of their predictions, which is large where the data are thin or contradictory.
The decisive test is leave-one-material-out generalization. Let be the set of materials in the database. For a material , let be the vector of its electro-optic targets and \vect{x}_m its inputs; let be the model trained on every record except those of , and the mean target over those same records. All of is removed from training, so the model has never seen that material, and its error is compared with a baseline that predicts the training mean:
where MAE is the mean absolute error. The test is run for every material with at least three electro-optic records.
The gold standard
A gate reports what it stopped, not what it let through. To measure the latter we built a gold standard: a set of records read by hand from a fixed set of papers, independent of any extraction pipeline, against which the pipeline’s output can be scored.
Choice of papers
Sixteen papers from the corpus were chosen to span the situations an extractor meets, rather than to be typical (tab:gold). They include the dense reference tensor sets for BaTiO, KNbO, and LiNbO, in which electro-optic, elastic, piezoelectric, dielectric, elasto-optic, and refractive constants share one table [37, 45, 49]; first-principles studies of BaTiO that print full Pockels matrices as a function of strain or lattice constant [40, 56]; measurements on bulk crystals of LiTaO, PIN–PMN–PT, KNbBO, and the zinc-blende semiconductors [71, 34, 50, 65, 63]; thin-film and device papers on BaTiO, BiFeO, SrTiO, and 3C-SiC, where a single effective value sits among many device figures [61, 69, 41, 60, 59]; and a review of potassium tantalate niobate [46].
Certification
Each value was read from the paper by hand and recorded with the page, the verbatim source text, the material, the tensor label as printed and its canonical form, the unit, the mechanical boundary condition (clamped or unclamped), the contribution type (electronic, ionic, piezoelectric, or total), the provenance, and the measurement conditions. Values the paper cites from other work were excluded, so that the gold standard, like the gate, contains each paper’s own results only. Some papers contribute very few gold values, and that is deliberate. The potassium tantalate niobate review reports a Pockels coefficient of about 2200 pm/V that is induced from the Kerr coefficient by a bias field [46], not an intrinsic linear coefficient, so it contributes no gold Pockels value; it tests whether a pipeline wrongly accepts one. The device papers similarly contribute one or two values each and many tempting numbers that are not coefficients.
Matching
A predicted record counts as correct only if its value, material, canonical component, boundary condition, and contribution type all match a gold record from the same paper; this is the full-record match used throughout. Numeric precision, which counts a number as correct regardless of its label, is reported separately because it isolates the errors that put the right number in the wrong slot.
| DDBand DDNavyMaterial | DDNavyKind of source | DDNavyReference |
| BaTiO | bulk reference tensor set (all families) | |
| KNbO | bulk reference tensor set (all families) | |
| LiNbO | bulk reference tensor set (all families) | |
| BaTiO | first principles, tetragonal, strain | |
| BaTiO | first principles, Pockels matrices | |
| BaTiO | strained thin film, cryogenic | |
| BaTiO | thin film and device, MHz to sub-THz | |
| LiTaO | bulk crystal, near infrared | |
| PIN--PMN--PT | single-domain bulk crystal | |
| KNbBO | new bulk crystal | |
| BiFeO | thin-film modulator | |
| SrTiO | thin film, cryogenic | |
| SiC (3C) | integrated modulator | |
| CdTe, ZnCdTe | bulk crystal, 1550nm | |
| GaP | bulk crystal, dispersion | |
| KTaNbO | review; field-induced response (control) |
Results
What the gate did in a production run
The most recent single production run (September 2026) processed the 49 PDFs of § 7. Three papers were withheld at detection as out of scope (hybrid or organic perovskites, or a multi-material survey), and 43 papers yielded at least one candidate value. The 816 candidates were divided into 414 accepted, 215 held for review, and 187 rejected, each held or rejected value with at least one typed reason (tab:tiers). The 414 accepted values come from 35 papers and 19 materials. They comprise 232 electro-optic values and 182 values of other families: refractive index (49), elastic (33), piezoelectric strain (30) and stress (11), elasto-optic (27), dielectric (16), optical dielectric (14), and density (2).
| DDBand DDNavyReason class | DDNavyHeld | DDNavyRejected | DDNavyTotal |
| Corrupted, missing, or unrecognised tensor label | 95 | 100 | 195 |
| Out-of-scope property family (e.g. phonon frequencies) | 0 | 137 | 137 |
| Cited or comparison value, not the paper's own result | 54 | 32 | 86 |
| Missing condition key (value from a series) | 27 | 0 | 27 |
| Family, label, or unit mismatch | 25 | 2 | 27 |
| Implausible magnitude for the stated context | 19 | 3 | 22 |
| Symmetry-forbidden component (zinc-blende rule) | 0 | 20 | 20 |
| Optical character recognition (OCR) corruption, or value not verbatim in the source | 5 | 12 | 17 |
| Not a measurement (equation number, imaginary mode) | 0 | 4 | 4 |
The provenance of the accepted electro-optic values matters for any model trained on them. Of the 232, 67 are tagged as experiment, 90 as first-principles calculation, 10 as part of a fitted self-consistent set, 4 as derived, and 61 as unspecified; 207 are intrinsic tensor components and 25 effective device coefficients. Only 50 carry complete temperature-and-strain context, 107 partial, and 75 none.
Measured extraction accuracy
We compared the production output with the certified gold standard of § 3.6, 303 values from 16 papers, using the full-record match.
These figures set the scope of everything that follows. The gate removes whole classes of error deterministically (tab:tiers,sec:caught), but about half of what it accepts still differs from the gold standard in at least one field. An audit of the accepted values from three dense legacy tensor papers found that 30 of 112 were defective, and that checks needing no reference data, chiefly applying symmetry rules to every crystal class rather than only to zinc-blende, would have caught 21 of them [23]. Precision-first is therefore a design priority of this pipeline and a direction of improvement, not an achieved level of accuracy.
The database used for modelling
The modelling results below use a consensus database that was assembled over several staged runs during the campaign, as the corpus, the material vocabulary, and the gate evolved. Its latest snapshot contains 448 consensus records over 22 inorganic materials from 39 contributing papers (tab:database). It spans perovskite ferroelectrics, niobates and tantalates, phosphate and iodate crystals, a lead-titanate relaxor family, and a still-thin semiconductor set, across electro-optic, refractive, dielectric, piezoelectric, and elastic families. Three papers that recurred under different file names were detected as byte-identical and de-duplicated.
| DDBand DDNavyMaterial | DDNavyRecords | DDNavyMaterial | DDNavyRecords |
| BaTiO | 149 | SrBaNbO(SBN) | 9 |
| BiFeO | 61 | AlN | 7 |
| LiNbO | 47 | CdTe | 6 |
| KNbO | 40 | Pb(Zr,Ti)O(PZT) | 6 |
| PMN--PT | 30 | SrTiO | 3 |
| LiIO | 20 | BaSrTiO | 2 |
| LiTaO | 19 | KTaNbO | 2 |
| KTiOPO(KTP) | 17 | KNbBO | 1 |
| PZN--PT | 13 | GaP | 1 |
| BiScO--BiGaO--PbTiO | 12 | InP | 1 |
| SiC | 1 | ||
| ZnTe | 1 |
Coverage drives out-of-distribution generalization
As the database came to cover more distinct materials, the surrogate’s ability to predict held-out materials grew with it. Table 1 records four snapshots taken during the campaign and Figure 8 plots them. The number of materials whose LOMO error beats their baseline (eq:lomo) rose as the number of materials rose , and the share of the majority material (BaTiO) fell from 56% to a third. The fraction of tested materials that generalize rose from 25% to 80% and then settled at 69%, because each new material is often a single-source, data-poor one that also joins the test pool.
The snapshots are not a controlled experiment. In a controlled experiment, one factor (here, the number of materials) would be varied while everything else was held fixed. Between our snapshots several things changed at once: the papers in the corpus, the material vocabulary, and the version of the gate. The rise in Table 1 is therefore consistent with coverage driving generalization, but it does not prove it, because a better gate or different papers could contribute to the same rise.
| Early | 367 | 10 | 56 | 1/4 | 0.600 | 0.603 |
| Merge | 517 | 15 | 41 | 6/8 | 0.450 | 0.640 |
| Rebuilt | 400 | 19 | 34 | 8/10 | 0.337 | 0.603 |
| Current | 448 | 22 | 33 | 9/13 | 0.348 | 0.608 |
Table 1. The coverage–generalization trajectory over four database snapshots. The LOMO denominator grows because each snapshot admits new, often data-poor materials into the test pool. CV MAE (cross-validated mean absolute error) and baseline are in units of the electro-optic coefficient. The snapshots differ in corpus, vocabulary, and gate version.
\begin{axis}[ddlabj,
xlabel={Materials in database}, ylabel={Held-out materials},
xmin=8, xmax=24, ymin=0, ymax=14, ytick={0,2,...,14},
legend pos=north west, legend cell align=left]
\addplot+[thick] coordinates {(10,1)(15,6)(19,8)(22,9)};
\addlegendentry{Generalize (LOMO beats baseline)}
\addplot+[thick, dashed] coordinates {(10,4)(15,8)(19,10)(22,13)};
\addlegendentry{Tested}
\end{axis}Figure 8. Out-of-distribution generalization tracks the number of distinct materials across the four snapshots of Table 1. The count of held-out materials that beat their baseline rises with the number of materials; the test pool (dashed) grows in step, which is why the fraction does not rise monotonically.
At the current snapshot the surrogate reaches a CV MAE of 0.35 units against a predict-the-mean baseline of 0.61, and nine of thirteen held-out materials beat their baseline (Table 2). They span perovskites (BaTiO, SrTiO), a ferrite (BiFeO), a niobate (SBN), a nitride (AlN), a tantalate (LiTaO), and the lead-titanate relaxor family (PMN–PT, PZN–PT, and the BiScO–BiGaO–PbTiO solid solution).
| LiTaO | 19 | 0.155 | 0.319 | Beat |
| SrTiO | 3 | 0.160 | 1.157 | Beat |
| PMN–PT | 4 | 0.239 | 0.867 | Beat |
| BiScO–BiGaO–PbTiO | 4 | 0.236 | 0.477 | Beat |
| SBN | 5 | 0.292 | 0.325 | Beat |
| PZN–PT | 8 | 0.520 | 0.791 | Beat |
| BaTiO | 92 | 0.534 | 0.755 | Beat |
| BiFeO | 48 | 0.810 | 1.188 | Beat |
| AlN | 6 | 0.846 | 1.577 | Beat |
| LiNbO | 18 | 0.341 | 0.293 | Miss |
| KTP | 11 | 0.340 | 0.224 | Miss |
| KNbO | 11 | 0.464 | 0.344 | Miss |
| CdTe | 3 | 0.846 | 0.420 | Miss |
Table 2. Leave-one-material-out generalization at the current snapshot. is the number of electro-optic records of the held-out material entering the LOMO fit (tab:database counts all families). means the held-out MAE is below the material’s own predict-the-mean baseline (eq:lomo). Errors are in units.
What the gate caught
Six interventions show the gate and the deterministic parsers acting on real extraction output. Each is a value that a recall-first pipeline would have entered wrong, mislabelled, or without indices.
Symmetry rejection. In the September run, a paper on GaInAsP/InP modulators [44] yielded , , , and for InP. InP is cubic zinc-blende, which permits only ((7)); all four were rejected (lst:gate), and the paper’s only accepted record is a refractive index.
A number look-alike. In an earlier staged run, a CdTe entry from a study of ZnCdTe crystals [65] read r_{41}=\SI{110}{pm/V}, far above any reported CdTe value. Tracing it to the source showed that “110” was the orientation along which the samples were cleaved. The record was removed; in the September run the same paper contributes the value it reports, r_{41}=\SI{15.5}{pm/V}.
Symbol recovery. The rule recovered a five-component tensor from the hydrothermal KTiOPO paper [48] that a literal reader would have discarded as unrecognised labels (§ 3.3).
Identity separation. A paper on 0.88 Pb(ZnNb)O–0.12 PbTiO [85] was, before the vocabulary was extended, attributed to PbTiO through its end-member formula. Adding PZN–PT as a distinct identity kept its large coefficient (r_{33}=\SI{134}{pm/V}) from being assigned to PbTiO.
Unit consistency. In the BaTiO reference tensor set [37], permittivities of about 4400 were read into elastic-constant slots. They were held instead of entering as elastic constants, because their source text carried no elastic unit.
Structured-tensor recovery. A first-principles study of BaTiO [56] printed its Pockels tensors as bracketed arrays of bare numbers, one per lattice constant. The deterministic parser assigned each entry its label from its position.
Limitations
Discussion and outlook
From a forward model to inverse design
The forward surrogate is not the end product; it is the operator that the framework’s generative stage inverts. (3) asks for the material configuration and operating conditions that maximise the electro-optic response. A model that generalizes across materials, and reports where it does not, turns that question into a constrained search:
where \vect{x}=(\vect{c},\vect{k}) joins the material configuration and the operating conditions of (3), is the space of candidates, the desired Pockels coefficient, a loss that is zero once the target is reached, the ensemble spread of § 3.5, which penalises extrapolation into data-poor regions with weight , a penalty with weight on charge imbalance, symmetry incompatibility, and synthesis infeasibility, and the domain over which the model has demonstrated LOMO generalization. The database enters this expression twice: it defines , and its per-material LOMO verdicts define . A candidate proposed outside that domain is flagged by rather than trusted.
What the inverse problem actually asks
For quantum and integrated photonics the target is rarely “the largest ”. A cryogenic modulator or transducer needs a large at a few kelvin, low optical loss, and a dielectric constant that does not make the electrodes impractically capacitive. The natural objective is therefore a figure of merit evaluated at the operating point, for example
where is the relative permittivity and () sets how strongly capacitance is penalised, which depends on the device. The database already records refractive index, dielectric constants, and electro-optic coefficients for the same materials, which is what (14) needs. Because the knobs \vect{k} are inputs of , the same model can be asked a second kind of question: not which new composition to make, but at which temperature, strain, or bias field an existing crystal performs best. For barium titanate on a chosen substrate, for example, the question becomes which strain state maximises (14) at 4 K, the regime that strained-film experiments are now probing [61, 60].
A closed loop
alg:loop states the intended loop. The surrogate proposes candidates inside ; first-principles calculations of the Pockels tensor, of the kind already in the corpus [56, 40, 68], screen them; the survivors are synthesised and measured; and every result, confirmed or not, returns to the database through the same gate, so that the next round of the model learns from it. The ensemble spread tells Stage 0 which material or condition to look for in the literature next [21].
What has to improve first
Three gaps separate the present model from this loop. Its resolution, about a factor of two (§ 4.6), is enough to rank candidates but not to promise a value. Only about a fifth of the electro-optic records carry complete temperature-and-strain context, so the condition dependence that (14) relies on is learned from few examples. And measured and computed values are mixed in one model. Each gap is addressed by data rather than by a different algorithm, which is why the next steps below concern the database.
Next steps
Scale the corpus, with acquisition held to the same standard
A first pass of Stage 0 assembled a catalogue of 1,032 records and 645 downloaded PDFs (560 unique). A screen with the pipeline’s own material detector and a search for coefficients reported in pm/V, which costs nothing to run, found that 57 of the 557 papers not already in the corpus report electro-optic coefficient values. Acquisition precision, not extraction cost, is therefore the first constraint on scale, and the coefficient screen will be applied as a gate before download.
Close the semiconductor gap
A few measured zinc-blende and wurtzite sources (ZnTe, GaAs, GaN) would give GaP and CdTe the chemical neighbours the composition features need.
Uncertainty-guided acquisition
The modelling stage already emits what an active-learning loop consumes: a per-material generalization verdict and a per-record confidence. Coupling the ensemble spread of eq:design back into the relevance ranker would let the system acquire the material, condition, or measurement that most reduces its own uncertainty [21].
Condition-resolved modelling
As coverage of conditions grows, the surrogate can move from a composition model toward the full r(\vect{c}, T, \varepsilon, E_{\mathrm{bias}}, \lambda) response that (3) targets. Cryogenic data are the priority, because they are the regime of quantum transduction and the one handbooks cover least.
Domain transfer
Nothing in the five-stage architecture is specific to electro-optics. The symmetry clause, the symbol-recovery rule, and the provenance weighting are instances of one pattern, encoding a domain’s physics as acceptance constraints, that transfers to battery, catalysis, and quantum-materials databases by substituting the vocabulary and the admissibility rules.
The end-to-end procedure
The pipeline of fig:pipeline is stated once as a procedure, so that the manuscript and the code agree on order and on where a value is allowed to stop.
The corpus
Table 3 lists the 48 distinct works in the 49-file corpus of the production run (one work appears both as a preprint and as its published version). Papers marked belong to the gold standard of § 3.6; the last three, on organic and hybrid materials, are outside the scope of the database.
| BaTiO | reference tensor set, bulk crystal | |
| KNbO | reference tensor set, bulk crystal | |
| LiNbO | reference tensor set, bulk crystal | |
| LiNbO(Er, In:Er doped) | ,of doped crystals | |
| LiTaO | near-infrared coefficients | |
| LiTaO(near-stoichiometric) | deep-ultraviolet electro-optics | |
| BaTiO | first principles, tetragonal phase | |
| BaTiO | first principles, Pockels matrices vs. lattice constant | |
| BaTiO | strained thin film, cryogenic | |
| BaTiO, LiNbO | thin film, MHz to sub-THz response | |
| BaTiO | thin film on silicon | |
| BaTiO | single-crystalline thin films | |
| BaTiO | thin-film tensor characterisation | |
| BaTiO | films on Si, thermal-rate control | |
| BaTiO(doped) | nanocrystals | |
| Lead-free perovskites | thin films, Pockels and Kerr | |
| BaSrTiO | electro-optic filter at 2 m | |
| SrTiO | thin film, cryogenic (preprint and published version) | |
| Perovskite oxides | design rules, first-principles screening | |
| BiFeO | thin-film optical modulators | |
| BiFeO | thin films, linear electro-optic effect | |
| PIN--PMN--PT | single-domain crystal | |
| Sm-doped PIN--PMN--PT | domain-engineered crystal | |
| Pb(Zr,Ti)O | plasmonic modulator, | |
| Pb(Zr,Ti)O | thin-film Pockels laser | |
| Pb(Zr,Ti)O(Ni-doped) | thin films | |
| KTaNbO | crystal growth and electro-optics (review) | |
| KTaNbO | microresonator tuning | |
| SrBaNbO | vs. temperature | |
| SrBaNbO | thin films (book chapter) | |
| KTiOPO | hydrothermal crystal, full tensor | |
| Na:KTiOPO, RbTiOAsO | low-frequency coefficients | |
| KNbBO | new electro-optic crystal | |
| -BaBO | Q-switch | |
| LiIO | first-principles optical properties | |
| KHPO(Mg-doped) | first-principles optical properties | |
| AlN | bulk nonlinear and electro-optic coefficients | |
| ScAlN | Pockels coefficient of a ferroelectric nitride | |
| SiC (3C) | integrated modulator | |
| GaAs/AlGaAs | ultralow-modulator | |
| GaAs | push-pull modulator, | |
| InP, GaInAsP | electro-optic modulators | |
| GaP | dispersion of | |
| ZnCdTe | single crystals at 1550 nm | |
| (Cd,Mg)Te | single crystals | |
| Metal-free perovskites | electro-optic modulation (out of scope) | |
| Hybrid halide perovskites | electro-optic modulation (out of scope) | |
| Conjugated molecules | organic electro-optics (out of scope) |
Table 3. The papers mined in this study.
References
- [1]Jain, A., Ong, S. P., Hautier, G., Chen, W., Richards, W. D., Dacek, S., Cholia, S., Gunter, D., Skinner, D., Ceder, G., & Persson, K. A. (2013). Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. APL Materials, 1(1), 011002. https://doi.org/10.1063/1.4812323
- [2]Saal, J. E., Kirklin, S., Aykol, M., Meredig, B., & Wolverton, C. (2013). Materials design and discovery with high-throughput density functional theory: The Open Quantum Materials Database (OQMD). JOM, 65(11), 1501–1509. https://doi.org/10.1007/s11837-013-0755-4
- [3]Curtarolo, S., Setyawan, W., Hart, G. L. W., Jahnatek, M., Chepulskii, R. V., Taylor, R. H., Wang, S., Xue, J., Yang, K., Levy, O., et al. (2012). AFLOW: An automatic framework for high-throughput materials discovery. Computational Materials Science, 58, 218–226. https://doi.org/10.1016/j.commatsci.2012.02.005
- [4]Belsky, A., Hellenbrandt, M., Karen, V. L., & Luksch, P. (2002). New developments in the Inorganic Crystal Structure Database (ICSD): accessibility in support of materials research and design. Acta Crystallographica Section B, 58(3), 364–369. https://doi.org/10.1107/S0108768102006948
- [5]Tshitoyan, V., Dagdelen, J., Weston, L., Dunn, A., Rong, Z., Kononova, O., Persson, K. A., Ceder, G., & Jain, A. (2019). Unsupervised word embeddings capture latent knowledge from materials science literature. Nature, 571(7763), 95–98. https://doi.org/10.1038/s41586-019-1335-8
- [6]Olivetti, E. A., Cole, J. M., Kim, E., Kononova, O., Ceder, G., Han, T. Y.-J., & Hiszpanski, A. M. (2020). Data-driven materials research enabled by natural language processing and information extraction. Applied Physics Reviews, 7(4), 041317. https://doi.org/10.1063/5.0021106
- [7]Kononova, O., He, T., Huo, H., Trewartha, A., Olivetti, E. A., & Ceder, G. (2021). Opportunities and challenges of text mining in materials research. iScience, 24(3), 102155. https://doi.org/10.1016/j.isci.2021.102155
- [8]Swain, M. C., & Cole, J. M. (2016). ChemDataExtractor: A toolkit for automated extraction of chemical information from the scientific literature. Journal of Chemical Information and Modeling, 56(10), 1894–1904. https://doi.org/10.1021/acs.jcim.6b00207
- [9]Court, C. J., & Cole, J. M. (2018). Auto-generated materials database of Curie and Néel temperatures via semi-supervised relationship extraction. Scientific Data, 5, 180111. https://doi.org/10.1038/sdata.2018.111
- [10]Huang, S., & Cole, J. M. (2020). A database of battery materials auto-generated using ChemDataExtractor. Scientific Data, 7, 260. https://doi.org/10.1038/s41597-020-00602-2
- [11]Kim, E., Huang, K., Saunders, A., McCallum, A., Ceder, G., & Olivetti, E. (2017). Materials synthesis insights from scientific literature via text extraction and machine learning. Chemistry of Materials, 29(21), 9436–9444. https://doi.org/10.1021/acs.chemmater.7b03500
- [12]Polak, M. P., & Morgan, D. (2024). Extracting accurate materials data from research papers with conversational language models and prompt engineering. Nature Communications, 15, 1569. https://doi.org/10.1038/s41467-024-45914-8
- [13]Dagdelen, J., Dunn, A., Lee, S., Walker, N., Rosen, A. S., Ceder, G., Persson, K. A., & Jain, A. (2024). Structured information extraction from scientific text with large language models. Nature Communications, 15, 1418. https://doi.org/10.1038/s41467-024-45563-x
- [14]Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 248. https://doi.org/10.1145/3571730
- [15]Nye, J. F. (1985). Physical Properties of Crystals: Their Representation by Tensors and Matrices. Oxford University Press.
- [16]Iglewicz, B., & Hoaglin, D. C. (1993). How to Detect and Handle Outliers (Vol. 16). ASQC Quality Press.
- [17]Leys, C., Ley, C., Klein, O., Bernard, P., & Licata, L. (2013). Detecting outliers: Do not use standard deviation around the mean, use absolute deviation around the median. Journal of Experimental Social Psychology, 49(4), 764–766. https://doi.org/10.1016/j.jesp.2013.03.013
- [18]Friedman, J. H. (2001). Greedy function approximation: A gradient boosting machine. The Annals of Statistics, 29(5), 1189–1232. https://doi.org/10.1214/aos/1013203451
- [19]Huber, P. J. (1964). Robust estimation of a location parameter. The Annals of Mathematical Statistics, 35(1), 73–101. https://doi.org/10.1214/aoms/1177703732
- [20]Northcutt, C. G., Jiang, L., & Chuang, I. L. (2021). Confident learning: Estimating uncertainty in dataset labels. Journal of Artificial Intelligence Research, 70, 1373–1411. https://doi.org/10.1613/jair.1.12125
- [21]Lookman, T., Balachandran, P. V., Xue, D., & Yuan, R. (2019). Active learning in materials science with emphasis on adaptive sampling using uncertainties for targeted design. Npj Computational Materials, 5, 21. https://doi.org/10.1038/s41524-019-0153-8
- [22]Datta, D. (2026). An autonomous scientific knowledge generation framework for AI-driven scientific discovery. In Preprint, Dibakar Datta Laboratory, New Jersey Institute of Technology. https://dibakardatta.net/dibakardatta.net
- [23]Datta, D. (2026). No value is better than a wrong value: Detecting, repairing, and tolerating erroneous data in literature-mined datasets for AI-driven science. In Preprint, Dibakar Datta Laboratory, New Jersey Institute of Technology. https://dibakardatta.net/dibakardatta.net
- [31]Abel, S., Eltes, F., Ortmann, J. E., Messner, A., Castera, P., Wagner, T., Urbonas, D., Rosa, A., Gutierrez, A. M., Tulli, D., Ma, P., Baeuerle, B., Josten, A., Heni, W., Caimi, D., et al. (2019). Large Pockels effect in micro- and nanostructured barium titanate integrated on silicon. Nature Materials, 18(1), 42–47. https://doi.org/10.1038/s41563-018-0208-0
- [34]Liu, X., Li, F., Li, J., Tan, P., Lei, J., Zhao, W., Song, K., Zheng, H., Tian, H., Li, F., Wei, X., & Xu, Z. (2024). Large Electro-Optic Coefficient in Single-Domain PIN-PMN-PT Single Crystal. Advanced Optical Materials, 12(33), 2400885. https://doi.org/10.1002/adom.202400885
- [37]Zgonik, M., Bernasconi, P., Duelli, M., Schlesser, R., Günter, P., Garrett, M. H., Rytz, D., Zhu, Y., & Wu, X. (1994). Dielectric, elastic, piezoelectric, electro-optic, and elasto-optic tensors of BaTiO$_3$ crystals. Physical Review B, 50(9), 5941–5949. https://doi.org/10.1103/PhysRevB.50.5941
- [40]Kim, I., Paoletta, T., & Demkov, A. A. (2023). Nature of electro-optic response in tetragonal BaTiO$_3$. Physical Review B, 108(11), 115201. https://doi.org/10.1103/PhysRevB.108.115201
- [41]Zhu, M., Du, Z., Liu, Q., Chen, B., Tsang, S. H., & Teo, E. H. T. (2016). Ferroelectric BiFeO$_3$ thin-film optical modulators. Applied Physics Letters, 108(23), 233502. https://doi.org/10.1063/1.4953201
- [44]Krauser, J., Albrecht, P., Bornholdt, C., Döldissen, W., Niggebrügge, U., Nolting, H.-P., & Schlak, M. (1986). Electro-optic modulators in GaInAsP/InP. Fiber Optic Sources and Detectors, 587, 180. https://doi.org/10.1117/12.951218
- [45]Zgonik, M., Schlesser, R., Biaggio, I., Voit, E., Tscherry, J., & Günter, P. (1993). Materials constants of KNbO$_3$ relevant for electro- and acousto-optics. Journal of Applied Physics, 74(2), 1287–1297. https://doi.org/10.1063/1.354934
- [46]Imai, T., Sasaura, M., Nakamura, K., & Fujiura, K. (2007). Crystal Growth and Electro-optic Properties of KTa$_1-x$Nb$_x$O$_3$. NTT Technical Review, 5(9), 4–11. https://doi.org/10.53829/ntr200709sp1
- [48]Zhou, H., He, X., Wu, W., Tong, J., Wang, J., Zuo, Y., Wu, Y., Zhang, C., & Hu, Z. (2023). Hydrothermal growth of KTiOPO$_4$ crystal for electro-optical application. Light: Science & Applications, 12(1), 23. https://doi.org/10.1038/s41377-022-01022-0
- [49]Jazbinšek, M., & Zgonik, M. (2002). Material tensor parameters of LiNbO$_3$ relevant for electro- and elasto-optics. Applied Physics B: Lasers and Optics, 74(4--5), 407–414. https://doi.org/10.1007/s003400200818
- [50]Lv, L., Yang, X., Wu, H., Su, R., Wang, Z., Su, B., Shan, P., Li, B., Long, X., Zhang, G., Xu, F., & He, C. (2025). A new-type electro-optic crystal: K$_3$Nb$_3$B$_2$O$_12$. Journal of Materiomics, 11(5), 100991. https://doi.org/10.1016/j.jmat.2024.100991
- [56]Paoletta, T., & Demkov, A. A. (2021). Pockels effect in low-temperature rhombohedral BaTiO$_3$. Physical Review B, 103(1), 014303. https://doi.org/10.1103/PhysRevB.103.014303
- [59]Powell, K., Li, L., Shams-Ansari, A., Wang, J., Meng, D., Sinclair, N., Deng, J., Lončar, M., & Yi, X. (2022). Integrated silicon carbide electro-optic modulator. Nature Communications, 13(1), 1851. https://doi.org/10.1038/s41467-022-29448-5
- [60]Ulrich, A., Brahim, K., Boelen, A., Debaets, M., Khalil, A., Sun, C., Huang, Y., Saseendran, S. S., Baryshnikova, M., Favia, P., Nuytten, T., Sergeant, S., Van Gasse, K., Kuyken, B., De Greve, K., et al. (2025). Engineering high Pockels coefficients in thin-film strontium titanate for cryogenic quantum electro-optic applications. Science, 390(6771), 390–393. https://doi.org/10.1126/science.adx3741
- [61]Suceava, A., Hazra, S., Ross, A., Philippi, I. R., Sotir, D., Brower, B., Ding, L., Zhu, Y., Zhang, Z., Sarkar, H., Sarker, S., Yang, Y., Sarker, S., Stoica, V. A., Schlom, D. G., et al. (2026). Colossal Cryogenic Electro-Optic Response Through Metastability in Strained BaTiO$_3$ Thin Films. Advanced Materials, 38(3), e07564. https://doi.org/10.1002/adma.202507564
- [63]Berozashvili, Y., Machavariani, S., Natsvlishvili, A., & Chirakadze, A. (1989). Dispersion of the linear electro-optic coefficients and the non-linear susceptibility in GaP. Journal of Physics D: Applied Physics, 22(5), 682–686. https://doi.org/10.1088/0022-3727/22/5/017
- [65]Jeon, Y., & Kang, H. S. (2007). Electro-Optic Coefficient Measurements for Zn$_x$Cd$_1-x$Te Single Crystals at 1550 nm Wavelength. Optical Review, 14(6), 373–375. https://doi.org/10.1007/s10043-007-0373-x
- [68]Hamze, A. K., Reynaud, M., Geler-Kremer, J., & Demkov, A. A. (2020). Design rules for strong electro-optic materials. Npj Computational Materials, 6(1), 130. https://doi.org/10.1038/s41524-020-00399-z
- [69]Chelladurai, D., Kohli, M., Winiger, J., Moor, D., Messner, A., Fedoryshyn, Y., Eleraky, M., Liu, Y., Wang, H., & Leuthold, J. (2025). Barium titanate and lithium niobate permittivity and Pockels coefficients from megahertz to sub-terahertz frequencies. Nature Materials, 24(6), 868–875. https://doi.org/10.1038/s41563-025-02158-1
- [71]Casson, J. L., Gahagan, K. T., Scrymgeour, D. A., Jain, R. K., Robinson, J. M., Gopalan, V., & Sander, R. K. (2004). Electro-optic coefficients of lithium tantalate at near-infrared wavelengths. Journal of the Optical Society of America B, 21(11), 1948. https://doi.org/10.1364/JOSAB.21.001948
- [72]Lauk, N., Sinclair, N., Barzanjeh, S., Covey, J. P., Saffman, M., Spiropulu, M., & Simon, C. (2020). Perspectives on quantum transduction. Quantum Science and Technology, 5(2), 020501. https://doi.org/10.1088/2058-9565/ab788a
- [73]Han, X., Fu, W., Zou, C.-L., Jiang, L., & Tang, H. X. (2021). Microwave-optical quantum frequency conversion. Optica, 8(8), 1050. https://doi.org/10.1364/OPTICA.425414
- [74]Wang, C., Zhang, M., Chen, X., Bertrand, M., Shams-Ansari, A., Chandrasekhar, S., Winzer, P., & Lončar, M. (2018). Integrated lithium niobate electro-optic modulators operating at CMOS-compatible voltages. Nature, 562(7725), 101–104. https://doi.org/10.1038/s41586-018-0551-y
- [75]Zhu, D., Shao, L., Yu, M., Cheng, R., Desiatov, B., Xin, C. J., Hu, Y., Holzgrafe, J., Ghosh, S., Shams-Ansari, A., Puma, E., Sinclair, N., Reimer, C., Zhang, M., & Lončar, M. (2021). Integrated photonics on thin-film lithium niobate. Advances in Optics and Photonics, 13(2), 242. https://doi.org/10.1364/AOP.411024
- [76]Zhang, M., Wang, C., Kharel, P., Zhu, D., & Lončar, M. (2021). Integrated lithium niobate electro-optic modulators: when performance meets scalability. Optica, 8(5), 652. https://doi.org/10.1364/OPTICA.415762
- [77]Eltes, F., Villarreal-Garcia, G. E., Caimi, D., Siegwart, H., Gentile, A. A., Hart, A., Stark, P., Marshall, G. D., Thompson, M. G., Barreto, J., Fompeyrine, J., & Abel, S. (2020). An integrated optical modulator operating at cryogenic temperatures. Nature Materials, 19(11), 1164–1168. https://doi.org/10.1038/s41563-020-0725-5
- [78]Sahu, R., Hease, W., Rueda, A., Arnold, G., Qiu, L., & Fink, J. M. (2022). Quantum-enabled operation of a microwave-optical interface. Nature Communications, 13, 1276. https://doi.org/10.1038/s41467-022-28924-2
- [79]Holzgrafe, J., Sinclair, N., Zhu, D., Shams-Ansari, A., Colangelo, M., Hu, Y., Zhang, M., Berggren, K. K., & Lončar, M. (2020). Cavity electro-optics in thin-film lithium niobate for efficient microwave-to-optical transduction. Optica, 7(12), 1714. https://doi.org/10.1364/OPTICA.397513
- [80]Youssefi, A., Shomroni, I., Joshi, Y. J., Bernier, N. R., Lukashchuk, A., Uhrich, P., Qiu, L., & Kippenberg, T. J. (2021). A cryogenic electro-optic interconnect for superconducting devices. Nature Electronics, 4(5), 326–332. https://doi.org/10.1038/s41928-021-00570-4
- [81]de Jong, M., Chen, W., Angsten, T., Jain, A., Notestine, R., Gamst, A., Sluiter, M., Ande, C. K., van der Zwaag, S., Plata, J. J., Toher, C., Curtarolo, S., Ceder, G., Persson, K. A., & Asta, M. (2015). Charting the complete elastic properties of inorganic crystalline compounds. Scientific Data, 2, 150009. https://doi.org/10.1038/sdata.2015.9
- [82]de Jong, M., Chen, W., Geerlings, H., Asta, M., & Persson, K. A. (2015). A database to enable discovery and design of piezoelectric materials. Scientific Data, 2, 150053. https://doi.org/10.1038/sdata.2015.53
- [83]Petousis, I., Mrdjenovich, D., Ballouz, E., Liu, M., Winston, D., Chen, W., Graf, T., Schladt, T. D., Persson, K. A., & Prinz, F. B. (2017). High-throughput screening of inorganic compounds for the discovery of novel dielectric and optical materials. Scientific Data, 4, 160134. https://doi.org/10.1038/sdata.2016.134
- [84]Basov, D. N., Averitt, R. D., & Hsieh, D. (2017). Towards properties on demand in quantum materials. Nature Materials, 16(11), 1077–1088. https://doi.org/10.1038/nmat5017
- [85]Lu, Y., Cheng, Z.-Y., Park, S.-E., Liu, S.-F., & Zhang, Q. (2000). Linear electro-optic effect of 0.88Pb(Zn$_1/3$Nb$_2/3$)O$_3$--0.12PbTiO$_3$ single crystal. Japanese Journal of Applied Physics, 39(1R), 141–145. https://doi.org/10.1143/JJAP.39.141
- [86]Boyd, R. W. (2020). Nonlinear Optics. Academic Press.
- [87]Yariv, A., & Yeh, P. (1984). Optical Waves in Crystals: Propagation and Control of Laser Radiation. Wiley.