Intelligent SPAD system and method

Abstract

The present invention relates to an intelligent Single-Photon Avalanche Diode (SPAD) system utilizing LEDs with reverse bias applied at or near the breakdown voltage, operating in Geiger mode. The system features an intelligent quenching mechanism, a control circuit with a programmable microcontroller, and embedded sensors such as thermistors. A learning system employing machine learning algorithms distinguishes true photon events from noise, enhancing detection accuracy. The intelligent quenching circuit dynamically adjusts the resistance to effectively quench photon-induced avalanches. The control circuit processes sensor data and pulse signals, using an inference engine to classify events and improve detection reliability. This invention offers a cost-effective, high-performance solution for precise photon detection, applicable in quantum computing, optical communication, and various scientific research fields.

Description

[0001] INTELLIGENT LED SPAD SYSTEM AND METHOD

[0002] FIELD

[0003] The present invention relates to photon detection systems and more specifically to a Single-Photon Avalanche Diode (SPAD) circuit utilizing an LED with reverse bias applied at or near the breakdown voltage. This invention integrates intelligent quenching mechanisms, embedded sensors such as thermistors, and programmable microcontrollers for enhanced photon detection, including a method training and deploying machine learning algorithms to improve the accuracy of distinguishing true photon events from anomalies. This invention is applicable in fields requiring precise photon detection, including quantum computing, optical communication, and various scientific research applications.

[0004] BACKGROUND

[0005] Single photon event detection is essential in various advanced technological applications, such as quantum computing, optical communication, and scientific research. Traditional SinglePhoton Avalanche Diodes (SPADs) are known for their high sensitivity and precision in detecting single photons. However, SPADs are expensive, require high operational voltages, and are sensitive to environmental conditions, affecting their performance [1], [2], Recent advancements have indicated that certain LEDs can be operated in Geiger mode, serving as cost- effective alternatives to SPADs, albeit with lower efficiency and susceptibility to noise phenomena, such as dark counts and after-pulsing [3], [4]. These inefficiencies hinder their practical utility in precise photon detection. The integration of machine learning algorithms presents a promising solution to these challenges. By employing machine learning techniques, it is possible to enhance the photon detection capabilities of LEDs in Geiger mode, distinguishing true photon events from noise. This invention leverages machine learning to optimize LEDbased SPAD circuits, thereby providing a viable and cost-effective alternative for accurate photon detection. The use of an LED as a SPAD and learning systems distinguishes this system from conventional SPAD designs, offering improved performance and cost efficiency [5]. SUMMARY

[0006] The present invention provides a Single-Photon Avalanche Diode (SPAD)-based photon detection system, featuring an intelligent quenching mechanism, noise correction learning models, and methods for training these models and the quenching system.

[0007] The disclosed embodiments include a SPAD detection device comprising an LED operating in Geiger mode through the application of a reverse bias voltage at or near the LED’s breakdown voltage. The system includes an intelligent quenching circuit composed of a discriminator, a programmable control circuit, an inference engine, a learning system, and embedded sensors.

[0008] In one exemplary embodiment, the photon detection system comprises a SPAD circuit with one or more LEDs reverse biased at or near their breakdown voltage, enabling a current pulse to flow whenever a photon induces an avalanche effect. The intelligent quenching circuit actively adjusts the resistance within the circuit to bring the LED below breakdown voltage, effectively quenching the avalanche. Embedded sensors, such as a thermistor, monitor environmental conditions and provide data to the inference engine via the control circuit. The control circuit, implemented as a microcontroller, hosts the inference engine. The inference engine uses a learning algorithm that leverages sensor data and pulse signals from the discriminator to classify pulses and distinguish true photon events from noise. The learning system includes a computer program that communicates with the microcontroller to receive sensor and pulse data, applying unsupervised learning and anomaly detection algorithms to identify noise, and uses this data to train and deploy the inference engines.

[0009] In one embodiment, the learning system is only connected to the SPAD circuit for the initial upload of the inference engine models.

[0010] In another embodiment, the inference engine is hosted externally from the control system.

[0011] Further objects, features, advantages, and properties of the system according to the present invention will become apparent from the detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In the following detailed portion of the present description, the teachings of the present application will be explained in more detail with reference to the example embodiments shown in the drawings, in which:

[0013] FIG 1A. Is a circuit diagram illustrating a prior art configuration of an LED-SPAD system as described by McCann [3].

[0014] FIG IB. Is a circuit diagram illustrating an improved LED-SPAD system with intelligent quenching system according to a first embodiment.

[0015] FIG 2. Is a circuit diagram illustrating the control circuit according to a first embodiment. FIG 3. Is a flowchart illustrating the inference engine program according to a first embodiment. FIG 4. Is a flowchart illustrating the learning system according to a first embodiment.

[0016] DETAILED DESCRIPTION

[0017] This detailed description presents embodiments of an intelligent Single-Photon Avalanche Diode (SPAD) system and associated learning method according to the teachings herein. While specific structures are described, it is to be understood that the teachings can be applied in various forms.

[0018] FIG. 1A. illustrates a conventional LED-based SPAD circuit described in McCann [3]. This setup includes a voltage source where the cathode of the LED is connected to the positive terminal and the anode is grounded via a resistor. This configuration is typical for detecting photon strikes, which trigger an Avalanche Effect in the LED's depletion zone, resulting in a current spike. This spike is detected by an operational amplifier (op-amp, or discriminator), where the inverting input is connected to the LED and the non-inverting input is grounded. The op-amp's output connects to a voltage divider, setting the output voltage (Vout), and is powered by +15V and +5V supplies.

[0019] FIG. IB illustrates an advanced configuration of the LED SPAD, incorporating elements from prior art. This configuration includes a voltage source 100, configured to apply a reverse bias voltage at the breakdown level to the cathode of LED 103. The circuit is powered by a 15V voltage regulator 101, which supplies power to certain components of the circuit, and a 5 V voltage regulator 102, which powers parts of the control circuit. The anode of LED 103 is connected to the collector of a transistor 201, which forms part of the control mechanism.

[0020] In alternative embodiments, the voltage source 100 may be a variable voltage source, allowing for the adjustment of the reverse bias voltage to optimize performance under various operating conditions or for specific photon detection requirements. Additionally, instead of fixed voltage regulators, adjustable voltage regulators could replace 15V regulator 101 and 5V regulator 102 to provide more precise control over the power supplied to the circuit components, thus accommodating variations in environmental conditions or operational requirements. In other embodiments, these voltage regulators can be replaced by alternate power sources.

[0021] Moreover, the use of LEDs with varying spectral sensitivities or breakdown voltages is contemplated, enabling the SPAD to be specialized for detecting specific types of photons, such as those in the ultraviolet or infrared spectra. A plurality of LEDs could also be employed in lieu of the single LED 103 to increase the detection area or to enable spatial resolution of photon strikes, thus enhancing the functionality and application scope of the device.

[0022] In this embodiment, resistor 200 is strategically connected between the anode of LED 103 and the ground. This connection serves the critical function of providing passive quenching, essential for stabilizing the circuit after photon detection events. Potentiometer 202, which is connected to the output of the 5V voltage regulator 102 and to the ground, facilitates adjustable voltage reference settings. Its output is concurrently routed to the analog inputs of the control system circuit via reference A2, enabling dynamic adjustments based on operational conditions.

[0023] The circuit integrates discriminator 203, An LM311 in this embodiment. 203 is configured with its non-inverting input grounded to ensure stability. Its inverting input taps into the junction between resistor 200 and transistor 201, optimizing signal sensitivity to changes induced by photon interactions. Resistor 204 forms a connection from the output of discriminator 203 to the +5V output of voltage regulator 102, thereby establishing a default high signal state. Additionally, another resistor 205 is similarly connected, enhancing the circuit's response to signal deviations.

[0024] When photon-induced events alter the voltage across resistor 200, diverging from the setpoint determined by potentiometer 204, the output of the op-amp discriminator 203 is effectively 'pulled-up' to the +5V supplied by regulator 102. This action triggers a signal that is transmitted to the control system circuit at input pin 302, where it is further processed to discern the nature of the event.

[0025] In alternative embodiments, the circuit architecture allows for the junction of transistor 201 and resistor 200 to be directly connected to the output of an adjustable +5V Voltage Regulator 102, allowing for real-time tuning of the bias voltage. This modification aims to refine the alignment of the non-inverted input signal with the reference voltage, thus enhancing detection accuracy. Additionally, to adapt to different operational requirements, transistor 201 could be replaced with a Field-Effect Transistor (FET) for improved power efficiency. For more precise thermal monitoring, a digital temperature sensor could replace thermistor 202, and in other embodiments wireless transmission capabilities could be added to enable remote data communication. In other embodiments, transistor 201 can be replaced by a Potentiometer or other components that would be obviously useful for rapidly controlling the current across the circuit.

[0026] Thermistor 202 is judiciously positioned in physical contact with LED 103 to monitor temperature variations that may affect the LED's performance. Depending on the specific application requirements, the thermistor may be secured with thermal glue or merely placed in close proximity to LED 103. It is powered by the +5V signal emanating from output pin 305 of the control system circuit, as delineated in FIG. 2. This arrangement ensures continuous monitoring of thermal conditions, which is vital for maintaining circuit integrity and performance under varying environmental conditions.

[0027] FIG. 2 depicts an embodiment of the control circuit system, wherein the system utilizes a microcontroller, exemplified by an Arduino Nano, to manage operations. The analog input pin 300 of the microcontroller is configured to receive voltage level readings from the junction of potentiometer 204 with the inverting input of discriminator 203. This setup captures not only voltage levels but also temperature readings from thermistor 202 via pin 304, and temporal data measuring the time between events in microseconds, triggered upon receiving an interrupt signal at pin 302 from the output of discriminator 203.

[0028] Pin 301 of the control circuit system outputs a "high" signal, approximately +5V, to maintain current flow through transistor 201. The system is designed with the capability to reduce the output of pin 301 to a low voltage, effectively ceasing current flow through transistor 201 and rapidly quenching the avalanche circuit. Additionally, the +5V output pin 305 is tasked with powering thermistor 202, thereby ensuring that the voltage levels remain within a safe operational range for the microcontroller.

[0029] Furthermore, the microcontroller includes a USB port 303, which serves multiple purposes: it can be used to upload the inference engine software from FIG. 3 onto the microcontroller, establish a connection with the external device hosting the learning system of FIG. 4, and supply power to the microcontroller.

[0030] In alternative embodiments, various microcontrollers besides an Arduino Nano might be utilized, each offering different capabilities in terms of processing power, memory capacity, and additional functionalities. For instance, advanced microcontrollers with integrated Wi-Fi or Bluetooth modules could be employed to enable wireless communication capabilities, facilitating remote monitoring and control of the SPAD system. Alternative voltage levels other than the +5V, such as the more modem +3.3V, may be used across the circuit instead.

[0031] Additionally, the use of a programmable logic controller (PLC) or a dedicated digital signal processor (DSP) could be considered for applications requiring enhanced processing speed or more complex computational tasks. Such alternatives would expand the utility and adaptability of the control circuit system across a broader range of photon detection applications and operational environments.

[0032] In FIG 3. the inference engine executes a continuous event loop 400 when loaded onto the chip. The event loop continues indefinitely until the interrupt 401 is triggered by a pulse on the control circuit system pin 302 from the discriminator of 203. The interrupt routine 402 then captures the analog (300 and 304) inputs and calculates the time in microseconds since the last pulse interrupt. In the exemplary embodiment, the system can be set in learning mode by flashing the appropriate program onto the device. In other embodiments, the learning mode can be triggered by an apparatus that sends a pulse onto an input pin of the Arduino, or another signal that would be obvious to a person having ordinary skill in the art. When the system is not in learning mode, the inference engine uses the uploaded model of 403 that was previously trained to use the combined inputs from pin 300 (Voltage at the Potentiometer 204), 304 (Thermistor 202 output) and the time since the last pulse to infer whether or not the pulse was caused by a true photon strike event or was due to noise. In the case the system detects noise, the routine at 404 sets the voltage across pin D3 301 “low”, stopping current across transistor 201 and immediately quenching the circuit. If a true photon event is detected, the routine at 405 will write the result to the target output. In the exemplary embodiment, the output will be written to the serial output of the USB 303. In alternative embodiments the output can be sent as a high or low signal via one of the microcontroller’s digital output pins. In the event the system is configured to operate in learning mode, the data from pins 300, 304, and the time since the last pulse will always be written to the learning system.

[0033] In alternate embodiments, the system can be in both learning mode and inference mode at the same time, writing all events to the learning system while performing inference on the events to determine if the intelligent quenching circuit should be activated.

[0034] In other embodiments, the inference is conducted on events from a SPAD LED circuit without performing any operation on the quenching system.

[0035] In another embodiment, the inference engine is updated in real time as better performing models become available.

[0036] In another embodiment, the inference engine is hosted on a device external to the control circuit system, and the pulse event information is sent to the device and the resulting inference is read from it.

[0037] In another embodiment, a plurality of LED SPAD circuits may be wired to a single inference engine that processes signal inputs across the multiple circuits and makes appropriate adjustments to the intelligent quenching system components depending on the collective input of all of the circuits.

[0038] FIG. 4 delineates the operations of the learning system. Routine 505 perpetuates in a continuous loop until a significant volume of samples has been collected from the control circuit system's routine 405, typically via USB connection 303 to a device operating the learning system. In the standard embodiment, this device is a computer system. Alternatively, the learning system may be hosted on a tablet, smartphone, or any other sufficiently powerful device capable of executing the intricate learning routine.

[0039] Upon amassing the requisite quantity of samples — exemplified herein by one million — the data undergoes processing in routine 501, which engineers and refines additional useful features for subsequent analysis. These engineered features encompass timestamps of the events, converted to seconds, intervals between events also expressed in seconds, and the cumulative elapsed time. Depending on the embodiment, additional or fewer features may be generated as deemed straightforward by those skilled in the art. The processed data is subjected to a goodness-of-fit evaluation in routine 502, assessing how closely the data's distribution conforms to a Poisson versus a Gaussian distribution. This step is critical for identifying true photon events, which typically align more closely with a Poisson distribution. Routine 502 may also generate statistical analyses such as Chi-squared or Komolgorov-Smimov tests, or even produce histograms that demonstrate the frequency of counts per second against observed and theoretical values. These diagnostic tools are vital for establishing a baseline understanding of the noisy data.

[0040] In routine 503, data clustering is performed specifically on the interval data using the elbow method to ascertain the optimal number of clusters. Routine 504 labels the cluster with the shortest average interval between events as likely after-pulses. Subsequently, routine 505 generates a modified dataset excluding these after-pulses, allowing the remaining data to be reclustered in routine 506 across multiple attributes — namely temperature, voltage from Potentiometer 204, and intervals between events — to determine the most informative clustering arrangement. Routine 507 labels clusters with elevated temperatures as "dark counts," and routine 508 labels the original dataset accordingly, with remaining data points classified as photons.

[0041] The entire dataset then progresses through an anomaly detection process in routine 509, utilizing an algorithm such as the Isolation Forest to distinguish anomalous from normal records. Following this, routine 510 assigns a label of T' for photons deemed normal by the algorithm, and 'O' for all others. Routine 511 employs advanced machine learning techniques to train models capable of accurately classifying true photon events from noise. Using the SKLearn Python library, the data is split into stratified training and test subsets, which are then applied to train models such as a RandomForestClassifier and a LogisticRegression classifier. These models are evaluated for their predictive accuracy against the test datasets through a 5 -fold cross-validation process. A DummyClassifier is also used to benchmark the models' performance by generating uniformly random predictions against the test set, and a student’s t-test assesses if the model outcomes are statistically superior to those of the dummy classifier.

[0042] Should the models underperform or the labeled data not exhibit improved fit to a Poisson distribution, routine 513 is enacted to apply deeper data science techniques to diagnose and rectify the underlying issues. Conversely, if the models demonstrate satisfactory performance, they may be deployed into the control circuit system in routine 514 as serialized model weights or through other methods familiar to those skilled in the art.

[0043] In alternative embodiments, the learning system incorporates an automated feedback mechanism that continuously updates the learning models as new data is ingested. This system dynamically adjusts the model parameters in real-time, employing advanced algorithms such as online machine learning techniques or incremental learning approaches. These enhancements enable the system to adapt to changes in data patterns or operational environments without manual reconfiguration, thereby improving the system’s responsiveness and accuracy over time.

[0044] Additionally, the learning system can employ real-time anomaly detection algorithms that actively learn from incoming data, allowing for immediate adjustments to model assessments based on the latest observations. This approach ensures that the system remains robust against evolving noise patterns and can handle large data streams efficiently.

[0045] These adaptive and continuous learning capabilities significantly extend the utility of the learning system, catering to applications where real-time data processing and adaptive learning are crucial, such as in dynamic photon detection environments or complex industrial settings.

[0046] Algorithmic Implementation of Inference Engine

[0047] / / Define pin constants

[0048] Define pulsePin as pin 2

[0049] Define analogPinAO as Analog Input pin AO

[0050] Define analogPinAl as Analog Input pin Al

[0051] / / Initialization Function

[0052] Function SETUP

[0053] Set pulsePin as INPUT

[0054] Set analogPinAO as INPUT

[0055] Set analogPinAl as INPUT

[0056] Initialize Serial communication at 115200 baud rate

[0057] Attach interrupt to pulsePin to call COUNT PULSES function on RISING edge

[0058] End Function / / Main Program Loop

[0059] Function LOOP

[0060] / / Do nothing

[0061] End Function

[0062] / / Function to handle pulse detection and calculate time differences

[0063] Function COUNT PULSES

[0064] Declare lastPulseTime as unsigned long and initialize to 0

[0065] Declare currentPulseTime as unsigned long

[0066] Declare timeDiff as unsigned long

[0067] / / Get current time in microseconds

[0068] Set currentPulseTime to current time in microseconds

[0069] / / Calculate time difference since last pulse

[0070] Set timeDiff to currentPulseTime minus lastPulseTime

[0071] / / Update last pulse time to current time

[0072] Set lastPulseTime to currentPulseTime

[0073] / / Read analog values from pins AO and Al

[0074] Declare analogValueAO as integer and set to analog value from analogPinAO

[0075] Declare analogValueAl as integer and set to analog value from analogPinAl

[0076] / / Output analog values and time difference via Serial

[0077] Print analogValueAO, analogValueAl, and timeDiff to Serial with commas separating values End Function

[0078] Algorithmic Implementation of Data Collection Routine

[0079] / / Define configuration constants

[0080] Define serial_port as 'C0M5' Define baud rate as 115200

[0081] Define output file as 'photons_light_24V.csv'

[0082] Define MAX COUNT as 1000000

[0083] / / Initialize counter

[0084] Declare count as integer and set to 0

[0085] / / Open the serial port with specified configuration

[0086] Open serial port at serial_port with baud rate and 1 -second timeout

[0087] Wait for 2 seconds to ensure serial connection is stable

[0088] / / Open the CSV file for writing

[0089] Open file at output file for writing as file

[0090] / / Create a CSV writer object for the file

[0091] Create CSV writer for file

[0092] / / Write header row to the CSV file

[0093] Write row ["Temperature", "Potentiometer", "TimeSinceLast"] to CSV using writer

[0094] / / Continuously read from the serial port

[0095] While true

[0096] / / Check if data is available in the serial input buffer

[0097] If serial input buffer is not empty

[0098] Read line from serial port and decode using UTF-8, ignoring errors, and strip whitespace

[0099] / / Ensure the line is not empty

[0100] If line is not empty

[0101] Split line by commas into data array

[0102] / / Check if data array has exactly three elements

[0103] If length of data array is 3

[0104] Write data array to CSV using writer

[0105] Increment count by 1 Print count to console

[0106] / / Break the loop if count reaches MAX_COUNT

[0107] If count is greater than or equal to MAX COUNT Break loop

[0108] Print "Logged {count} records to {output file}"

[0109] / / Close the serial port

[0110] Close serial port

[0111] Algorithmic Implementation of Learning System

[0112] / / Define necessary imports

[0113] Include libraries for data manipulation (Pandas), numerical operations (NumPy), and visualization (Matplotlib)

[0114] / / Read data from a CSV file into a DataFrame

[0115] Read CSV file named "photons_light_24V.csv" into DataFrame named 'photons_light_24v'

[0116] / / Display the first few rows of the DataFrame

[0117] Display head of 'photons_light_24v'

[0118] / / Data Transformation

[0119] Add a column 'Timestamp' to 'photons_light_24v' by taking the cumulative sum of 'TimeSinceLast'

[0120] Convert 'Timestamp' from microseconds to seconds and store in a new column

[0121] 'T imestamp seconds'

[0122] Calculate the time between events in seconds and store in 'TimeBetweenEvents seconds' Calculate elapsed time from the start of the dataset and store in 'Elapsed seconds'

[0123] / / Display the first few rows of the DataFrame to verify new columns

[0124] Display head of 'photons_light_24v'

[0125] / / Generate descriptive statistics of the DataFrame

[0126] Display descriptive statistics of 'photons_light_24v'

[0127] / / Data Visualization

[0128] Plot 'Temperature' against 'Timestamp seconds' using a line plot

[0129] / / Data Cleaning

[0130] Filter data to include only records where 'Timestamp seconds' is greater than 100 seconds

[0131] / / Replot Temperature over Time with cleaned data

[0132] Plot 'Temperature' against 'Timestamp seconds' from cleaned data using a line plot

[0133] / / Plot Potentiometer over Time with cleaned data

[0134] Plot 'Potentiometer' against 'Timestamp seconds' from cleaned data using a line plot

[0135] / / Analyze and fit distributions to the time between events

[0136] For each specified bin width [0.015, 0.15, 1.5 seconds]

[0137] Call function 'plot histograms and fits' with cleaned data and current bin width / / Clustering

[0138] Define a function 'cluster events' that performs clustering on given columns of a DataFrame

[0139] Cluster data based on 'TimeBetweenEvents seconds'

[0140] Visualize the clustering results

[0141] / / Labeling clusters and handling labeled data

[0142] Label clusters as 'Afterpulse' or 'Photon' based on clustering results

[0143] Describe data labeled as 'Afterpulse'

[0144] / / Further data analysis on no afterpulse data

[0145] Remove 'Afterpulse' labeled data and describe the filtered dataset

[0146] Recalculate timestamps for filtered data and describe

[0147] For each bin width [0.015, 0.15, 1.5 seconds]

[0148] Call function 'plot histograms and fits' with no afterpulse data and current bin width

[0149] / / Isolation Forest Anomaly Detection

[0150] Define a function to find the optimal contamination level for Isolation Forest

[0151] Run Isolation Forest on labeled data and visualize the results for normal and anomalous points

[0152] / / Model Training and Evaluation

[0153] Split data into training and testing sets

[0154] Train and evaluate Random Forest and Logistic Regression models Compare models to a Dummy Classifier using T-tests

[0155] / / Serialization

[0156] Serialize and save the trained models to files for future use

[0157] REFERENCES CITED

[0158] 1] F. Zappa, M. Ghioni, S. Cova, C. Samori, and A. C. Giudice, "An Integrated Active- Quenching Circuit for Single-Photon Avalanche Diodes," IEEE Transactions on Instrumentation and Measurement, vol. 49, no. 6, p. 1167, Dec. 2000.

[0159] 2] J. Newport, "LEDs as Single-Photon Avalanche Photodiodes," American University, Washington, D.C., 2018.

[0160] 3] L. I. McCann, "Introducing students to single photon detection with a reverse-biased LED in avalanche mode," University of Wisconsin - River Falls, Physics Department, 410 S. 3rd St., River Falls, WI, 54022, 2015.

[0161] 4] R. A. Bianchi, M. M. Vignetti, and B. Rae, "Quenching of a SPAD," U.S. Patent Application US 20210105427 Al, Apr. 08, 2021.

[0162] 5] R. K. Henderson and J. Richardson, "Single Photon Avalanche Diodes," U.S. Patent 8,217,436 B2, Jul. 10, 2012.

Claims

AMENDED CLAIMS received by the International Bureau on 12 October 2025 (12.10.2025)1. An intelligent single-photon avalanche diode (SPAD) system comprising: at least one light-emitting diode (LED) configured to operate in Geiger mode by application of a reverse-bias voltage at or near a breakdown voltage of a cathode of the LED; a quenching circuit coupled to the LED; a microcontroller-based control circuit interfaced to a discriminator configured to output a pulse in response to photon events and to one or more sensors including a thermistor; and a learning system executing an inference engine; wherein the control circuit is configured to:(i) in response to the pulse, capture event features including at least inter-pulse time, temperature, and bias conditions;(ii) classify the event with the inference engine to determine a presence of a true photon event or noise;(iii) adjust, in real time and based on the classification, at least one operating parameter of the photon-detection circuit including the reverse-bias voltage or a quench-control parameter, thereby implementing a closed feedback loop that suppresses noise-induced avalanches while maintaining Geiger-mode operation; and(iv) output the classification to an external device.

2. The system of claim 1, further comprising a voltage source configured to apply the reversebias voltage to a cathode of the LED.

3. The system of claim 2, wherein the voltage source is variable to allow adjustment of the reverse-bias voltage.

4. The system of claim 1, comprising a plurality of LEDs configured to operate in Geiger mode.

5. The system of claim 1, wherein the microcontroller is configured to receive readings from the one or more sensors, compute time intervals between detected photon events, and actuate the quenching circuit to allow continuation or to quench an avalanche based on the classification.

6. The system of claim 1, further comprising a thermistor in thermal contact with at least one LED, wherein the microcontroller processes outputs from the thermistor to adjust operational parameters based on detected temperature variations.

7. The system of claim 1, wherein the microcontroller is further configured to communicate wired or wirelessly with an external device to transmit data and receive control instructions.

8. The system of claim 1, wherein the control circuit is configured to adjust a discriminator threshold based on the classification.

9. The system of claim 1, wherein capturing event features is interrupt-driven.

10. A method for detecting photon events using the system of claim 1, comprising: applying a reverse-bias voltage to at least one LED to operate it in Geiger mode; employing one or more sensors to capture data related to photon strikes; triggering a capture routine in response to a photon pulse to acquire event features; analyzing the captured data with the inference engine to determine a presence of true photon events or noise; adjusting at least one operating parameter of the photon-detection circuit based on the analysis to quench an avalanche when noise is detected while maintaining Geiger mode operation; and outputting the results of the photon event analysis to an external device.

11. The method of claim 10, further comprising continuously updating the inference model with new data to improve detection accuracy over time.

12. The method of claim 10, wherein real-time data processing and model updates are performed using online machine-learning techniques by the microcontroller.

13. The method of claim 11 or 12, further comprising wirelessly transmitting the processed data and analysis results to an external device for monitoring or further analysis.[0001][0002]STATEMENT UNDER ARTICLE 19 (1 )[0003]The replacement claims clarify that the invention concerns an intelligent single-photon avalanche diode (SPAD) system in which a learning system executes an inference engine whose outputs directly govern, in real time, operating conditions of an LED operated in Geiger mode. A microcontroller-based control circuit receives pulses from a discriminator and inputs from one or more sensors including a thermistor. For each detected event, the control circuit captures features such as inter-pulse time, temperature and bias conditions, classifies the event with the inference engine as a true photon event or noise, and immediately adjusts at least one operating parameter of the photon-detection circuit, including reverse-bias voltage and / or a quench-control parameter. This implements a closed feedback loop that suppresses noise-induced avalanches while maintaining Geiger-mode operation.[0004]The clarified claim language expresses the event-driven capture —> inference —> control update sequence that is central to system behavior. By tying classification outputs to real-time adjustments of bias and quench parameters, the system reduces false positives and afterpulsing, stabilizes performance across temperature and bias drift, and improves dynamic range and count-rate stability. The system further provides an output of the classification to an external device for monitoring or downstream processing.[0005]Method claims recite corresponding steps that apply a reverse-bias to operate the LED in Geiger mode, acquire event features in response to a photon pulse, analyze the features using the inference engine to determine the presence of true photon events or noise, adjust operating parameters based on the analysis to maintain stable Geiger-mode operation, and output results. Dependent claims set forth optional features including interrupt-driven capture, model updates using online machine-learning techniques, wired or wireless communication with an external device, and device-level variations such as multiple LEDs.[0006]No subject matter has been added; the amendments clarify the structure and operation of the learning-guided closed loop implemented in the disclosed SPAD system.

Images