#4944 Optical Matrix Computing
#4944 Optical Matrix Computing
#4944 For the following goal, would acrylic sheets or glass sheets work best? """Develop an optical computing component built of a stack of etched acrylic sheets where the etching is the program and light passes from top of stack to bottom with solution projected below."""
For an etched diffractive optical computing stack (D^2NN), high-optical-grade cell-cast acrylic (PMMA) works best for rapid prototyping and fabrication, but optical-grade glass (borosilicate/fused silica) is superior if absolute wavefront coherence, dimensional rigidity, and sub-micron phase control are required.
For most benchtop optical computing builds, cell-cast acrylic is the practical choice, though each material introduces specific physical trade-offs in an optical processing pipeline.
Key Trade-Off Comparison
| Metric / Parameter | Cell-Cast Acrylic (PMMA) | Optical Glass (Borosilicate / Fused Silica) | Impact on Optical Computing |
| Machinability / Etch Control | Excellent. Vaporizes cleanly under UV or short-pulse lasers with high material removal rates. | Difficult. Requires chemical wet etching (HF), precision abrasive masking, or specialized femtosecond lasers. | Direct impact on fabrication iteration speed and feature geometry. |
| Surface Figure & Flatness | Moderate. Prone to minor warpage, sag across stacked layers, and thickness variation (+/- 5--10%). | Exceptional. Flatness down to lambda/4 or lambda/10 across large apertures. | Stacked cumulative phase error: thickness variation alters optical path length (OPL = n cdot d). |
| Internal Stress & Birefringence | Present. Mechanical stress or localized heating introduces stress-induced birefringence (Delta n). | Near Zero. Homogeneous isotropic index (n ~ 1.46--1.52 depending on substrate). | Birefringence causes polarization rotation, phase noise, and crosstalk between diffractive layers. |
| Thermal Sensitivity ($dn/dT$) | High. Expansion coefficient ~ 70 x 10^{-6} /K; dn/dT is relatively high and negative. | Extremely Low. Borosilicate/Silica have near-zero thermal expansion and drift. | High-intensity illumination can cause local thermal lensing, drifting the computed output matrix. |
| Optical Transmission | High (>92%). Minimal bulk scattering across visible and near-IR bands. | Superior (>95--99%). Low absorption, zero UV degradation or solarization. | Throughput efficiency across a stack of 4–10+ consecutive layers. |
Critical Engineering Considerations for the Stack
Phase Modulation vs. Scatter Modulation
Optical neural networks (e.g., Lin et al., Science 2018) encode matrix weights as phase shifts Delta phi = {2pi} / {lambda}(n - 1) Delta h, requiring precise, multi-depth or continuously variable relief etches.
Laser-etching standard acrylic often produces micro-fractured, diffusing surfaces (white frosty scatter) rather than smooth optical phase profiles. If your model relies on coherent phase modulation, glass with chemical lithography or specialized cast-monomer grayscale UV ablation yields significantly smoother phase surfaces.
If your computing scheme relies on amplitude masks or binary blocking grids, acrylic etched with a dark infill (or thin metallic film ablation on acrylic) works reliably.
Cumulative Optical Path Length (OPL) Error
In a multi-pane stack where light propagates sequentially:
Delta Phi_total = sum_{k=1}^{N} {2pi} / {lambda} [n_k cdot d_k + (n_air) cdot z_gap]Standard commercial cast acrylic has sheet thickness tolerances of +/- 5--10%. Across a 5-layer stack, a 100 mu m variance in sheet thickness corresponds to hundreds of wavelengths of unmodeled phase drift, destroying coherent wave interference unless every pane is individually mapped and compensated for in the training algorithm.
Substrate Recommendation
Rapid Prototyping / Amplitude / Coarse Phase: High-grade optical cell-cast acrylic (GS). It does not melt into gummy beads like extruded acrylic, cuts cleanly, and allows rapid iteration of layer layouts.
Coherent Diffractive Deep Neural Networks (D^2NN): Borosilicate float glass (such as Schott Borofloat 33) or fused silica. The structural rigidity, low internal stress, tight thickness tolerances, and thermal stability ensure the physical hardware accurately computes the mathematically modeled diffraction pattern.
From <https://gemini.google.com/app/83255bb8496a1242> Google Gemini (3.8 Flash)
Given a stack of glass panes etched grid of 16 cells each with specific etching (TBD), what proof-of-concept optical computing could be done with this stack?
A stack of etched glass panes arranged in a 16-cell (4 x 4) grid implements a discretized Diffractive Optical Neural Network (D^2NN) or a cascade of complex spatial phase plates.
Because optical diffraction allows light emanating from any single cell on Pane N to spread and overlap onto multiple (or all) cells on Pane N+1, each inter-pane free-space gap acts as an analog all-to-all optical interconnect matrix, while the etched depths introduce controlled phase delays (Delta phi = 2pi / {lambda}(n - 1) Delta d).
Here are the most viable proof-of-concept demonstrations you can execute with this physical configuration:
1. Non-Separable Boolean Logic Engine (XOR / Full Adder)
Concept: Linear classifiers cannot solve XOR without hidden layers or non-linear mapping. In diffractive optics, interference performs linear matrix-vector transformations across layers, and thresholding at the detector (photodiode or camera) provides the necessary non-linear activation.
Input Encoding: 2 to 4 cells on Pane 1 represent input bits (e.g., Cell A illuminated = 1, dark = 0; or binary phase {0, pi).
Operation: 2–3 intermediate etched panes route and interfere the wavefronts destructively or constructively.
Readout: Output pane maps result to target detector quadrants (e.g., top right = 1, bottom right = 0).
POC Value: Proves multi-layer optical routing solves linearly non-separable problems using passive interference.
2. Low-Resolution Pattern Classifier (Toy "MNIST" / 4-Pixel Glyphs)
Concept: Optical classification of binary characters (e.g., 2 x 2 pixel patterns or cross/horizontal/vertical strokes) mapped across the input grid.
Operation:
Input is a beam mask encoding characters:
+,X,-, or|.The etched pane stack applies a fixed complex weight matrix W = P_k cdot D_{k-1} cdot P_{k-1} ... D_1 cdot P_1, where P_i are diagonal phase-shift matrices and D_i are free-space diffraction propagation operators (Rayleigh-Sommerfeld/Fresnel kernels).
Readout: Output focus directly lands all energy onto a designated target cell corresponding to class 0, 1, 2, or 3.
POC Value: Demonstrates zero-latency inference—classification occurs entirely at the speed of light during propagation.
3. All-Optical 4-Channel Spatial Router / Crossbar Switch
Concept: Passive spatial demultiplexing or unitary beam sorting.
Operation: An incident beam on input cell (i, j) is steered deterministically to a designated output cell (x, y) on the final pane depending on wavelength, input angle, or spatial mode.
POC Value: Serves as a passive interconnect switch or permutation matrix (16 x 16 permutation space), demonstrating routing without electro-optic conversion.
4. 2D Discrete Optical Transform (Hadamard / Fourier Pre-processor)
Concept: Emulating an optical 4-point or 16-point discrete spatial transform.
Operation: Pane depths are calculated so the overall cascade approximates a 16 x 16 discrete Fourier transform (DFT) or Walsh-Hadamard transform matrix.
Readout: Spatial frequencies appear directly separated at the output focal plane.
POC Value: Demonstrates analog mathematical preprocessing for edge detection, compression, or spatial filtering prior to digital capture.
Engineering Parameters for Design
| Parameter | Constraint / Rule of Thumb |
| Etch Depth Step (Delta d) | Delta d = {lambda} / {2(n_glass - 1)} for a pi-phase shift (~ 532 nm for green laser with n ~ 1.5) |
| Inter-Pane Distance (z) | Must satisfy the Fresnel diffraction regime: z ~ {w_cell^2} / {lambda}, ensuring the beam from one cell expands enough to cover neighboring cells |
| Phase Quantization | Binary phase (0, pi) via 1-step etch, or 4-level (0, pi / 2, pi, 3pi / 2) via dual-mask etching |
Can the stack with 1024 x 1024 cell matrices be encoded to classify particular objects in it line of vision?
Yes. Scaling from a 4 x 4 grid to a 1024 x 1024 matrix per pane elevates the stack from a toy demonstration into a full-scale Diffractive Deep Neural Network (D^2NN) capable of real-world object classification.
At 1024 x 1024, each glass layer contains approximately 1.05 million trainable phase neurons (1,048,576 individual etched features). A stack of 3 to 5 panes provides between 3.1 and 5.2 million optical parameters, matching or exceeding the parameter count of standard edge vision models like MobileNetV2 or SqueezeNet.
How the Optical Classification Works
Scene Input (Line of Vision):
Coherent light (e.g., from an expanded laser beam illuminating the target scene, or through a narrowband filter) bounces off or transmits through an object in the field of view.
An objective lens images the target scene onto the front face of Pane 1, establishing the complex spatial field U_0(x, y).
Wavefront Modulation & Interconnection:
Phase Delay (Weights): As the wavefront passes through cell (i, j) on a pane, the etched depth introduces a local phase delay Delta phi(i, j) = {2pi} / {lambda}(n - 1) Delta d.
Diffractive Interconnection (Synapses): In the free-space gap between Pane k and Pane k+1, light diffracts according to the Huygens-Fresnel principle / Rayleigh-Sommerfeld diffraction. Every single cell on Pane k broadcasts an expanding secondary wave that overlaps and interferes with thousands of neighboring cells on Pane k+1. This analog interference acts as an all-to-all matrix multiplication performed in flight.
Classification Readout (Zero Power, Zero Latency):
At the back focal plane (Pane N or a bare sensor array), the area is partitioned into predefined detection regions—one for each target class (e.g., Region 0 = Pedestrian, Region 1 = Vehicle, Region 2 = Stop Sign).
The layer etch patterns are mathematically optimized so that when a specific object enters the line of sight, destructive interference cancels light across incorrect bins, and constructive interference focuses high optical intensity exclusively into the correct class detector.
The "argmax" operation is completed passively at the speed of light (sim 1 ns for a 30 cm stack), with square-law intensity detection (I = vert{}E vert{}^2) providing natural rectification.
Key Capabilities vs. Physical Constraints at 1024 x 1024
| Aspect | Operational Reality |
| Model Capacity | Easily handles 10-class to 100-class object recognition (e.g., full MNIST, Fashion-MNIST, CIFAR-10 coarse classes, or silhouette identification). |
| Etch Feature Size | For a 50 mm x 50 mm optic, each cell is ~ 48.8 mu m. For a 25 mm substrate, pitch shrinks to ~ 24.4 mu m. Both are well within standard photolithography or multi-level reactive ion etching (RIE) limits. |
| Diffraction Angle (theta) | Governed by sin theta ~ {lambda} / {p} (where p is cell pitch). Larger features yield smaller diffraction angles, requiring precise tuning of inter-pane spacing (z) so adjacent receptive fields properly overlap. |
| Invariance | Scale, translation, and rotational shifts must be baked directly into the training dataset during optical backpropagation; otherwise, moving an object across the field of view shifts the interference pattern away from target detectors. |
Training and Encoding Pipeline
Training is done digitally before fabricating the physical glass:
Forward Pass (Digital Simulation): Calculate wave propagation through each layer using the Angular Spectrum Method (ASM):
U_{k+1} = mathcal{F}^{-1}{ mathcal{F} {U_k cdot e^{i phi_k}\} cdot H_z }where H_z(f_x, f_y) = e^{i {2pi} / {lambda} z sqrt{1 - (lambda f_x)^2 - (lambda f_y)^2}} is the spatial frequency transfer function of free space.
Loss & Backpropagation: Compute cross-entropy loss between detector energy readouts and ground truth labels. Use gradient descent (Adam optimizer) to compute {partial mathcal{L}} /.{partial phi_{k}(i, j)} across all 1024 x 1024 pixels per layer.
Quantization & GDSII Generation: Discretize the learned continuous phase values into etch masks (e.g., 8 to 16 etch steps using 3 to 4 binary masks, or continuous grayscale laser writing) to etch the glass substrate.
Watching physical setups of diffractive neural networks in action illustrates the real-world alignment tolerances and beam paths required:
Rather than neural network, how about classifying images via etching patters across the stack?
Classifying scenes or objects across etched glass panes without neural network weights relies on deterministic physical optics—primarily matched spatial filtering, interferometric nulling, and optical transformation matrices.
Instead of treating cells as gradient-tuned weights, the etched relief profiles directly implement closed-form mathematical operators (Fourier transforms, matched phase templates, or coordinate transformations) that collapse specific optical wavefronts into high-contrast detection peaks.
1. Cascaded VanderLugt / Matched Phase Filter
Physics: A classical optical correlator performs 2D cross-correlation via the Fourier domain:
g(x, y) = f(x, y) star h(x, y) <----> G(u, v) = F(u, v) cdot H^*(u, v)Etched Implementation:
Pane 1 & Lens: A diffractive lens profile etched across Pane 1 projects the optical Fourier transform of the incoming object onto Pane 2.
Pane 2 (Matched Phase Conjugate): Etched with the exact conjugate phase spectrum e^{-i Phi_{target}(u, v)} of the target object. When the specific object appears in the field of view, its wavefront phase is flattened into a uniform planar wave (Phi - Phi = 0). Any non-matching object leaves a chaotic, scrambled wavefront.
Pane 3 (Inverse Transform / Concentrator): An etched Fresnel zone plate or reverse Fourier element focuses the flattened planar wave into a diffraction-limited correlation spike (Bragg spot) on an axial detector.
Classification Output: If the correlation peak intensity exceeds a calibrated threshold (I > I_thresh), the specific object is present.
2. Multi-Channel Spatial Multiplexing (Synthetic Discriminant Functions)
Physics: To detect multiple classes or handle angular/scale variations without retraining, the panes implement a bank of orthogonal spatial filters via Synthetic Discriminant Functions (SDF).
Etched Implementation:
Pane 1 splits the incident scene into M directional diffraction orders (using an etched blazed/Dammann grating).
Pane 2 hosts a mosaic of deterministic phase filters: Sub-region A matches "Class 1", Sub-region B matches "Class 2", Sub-region C matches "Class 3".
Pane 3 steers each correlated sub-beam to a dedicated spatial bucket or pinhole photodiode.
Classification Output: A direct 1-of-N spatial energy readout without software inference.
3. Destructive Nulling / Optical Parity Checking
Physics: Instead of searching for bright peaks, this approach uses phase inversion to cancel light from known targets.
Etched Implementation:
The stack is etched to act as a destructive interference nuller (similar to a stellar coronagraph).
For the target reference silhouette or spatial mode, the internal path lengths across the stack introduce exact pi phase offsets between opposing spatial halves of the wavefront.
Classification Output: When the target is present and aligned, the central detector reads zero intensity (complete dark null). Any foreign object, aberration, or intruder breaks symmetry and causes an immediate bright flash on the photodetector.
4. Deterministic Coordinate Transformation (Log-Polar / Mellin Transform)
Physics: The greatest weakness of raw optical matching is sensitivity to rotation and scale. A passive optical stack can deterministically eliminate this via geometric phase transformations.
Etched Implementation:
Pane 1 (Conformal Mapper): Etched with a calculated phase profile Phi(x, y) that geometrically remaps Cartesian coordinates (x, y) into log-polar coordinates (ln r, theta).
Pane 2 (Phase Corrector): Cancels distortion and collimates the transformed beam.
In log-polar space, object rotation becomes a simple linear shift along the theta-axis, and scale change becomes a linear shift along the ln r-axis.
Pane 3 (Fourier Transform): An optical Fourier transform across the log-polar beam produces a Mellin transform that is completely invariant to both scale and rotation.
Classification Output: A downstream static matched filter then identifies the object regardless of whether it is rotated or moved closer/further from the optic.
Deterministic Optics vs. Diffractive Neural Networks (D^2NN)
| Parameter | Classical Etched Optics (Matched Filters / Transforms) | Diffractive Neural Network (D2NN) |
| Design Method | Analytical equations, Fourier optics, wave-optics beam propagation (BPM) | Empirical optimization (gradient descent / Adam backpropagation) |
| Generalization | Rigid: Identifies mathematically exact matches or spatial harmonics | Flexible: Interpolates across noisy, distorted, or handwritten variations |
| Intra-class Variance | Low tolerance for intra-class variation unless SDFs or conformal mappings are used | High tolerance (absorbs natural intra-class variance through training) |
| Interpretability | 100% deterministic: Every etched contour maps to a known Fourier component or phase shift | Black-box: Phase surfaces resemble pseudo-random speckle patterns |
| Tolerancing | Precise phase matching required at the design wavelength (Delta phi propto lambda^{-1}) | Moderate tolerance; network can be trained with simulated fabrication errors |
From <https://gemini.google.com/app/4307c4bc4fd4fe19> Google Gemini (3.8 Flash)
Comments
Post a Comment