A method and system for coherent beam combination

Abstract

Coherent beam combination (CBC) of a tiled laser array 3, includes directing a combined beam derived from multiple optical outputs at a target 11. A light sensor 4 to capture an optical intensity valu

Description

Coherent Beam Combination (CBC), where multiple lasers are combined into a single brighter beam with reduced divergence, is an essential technique for the power scaling of high power laser systems. CBC systems usually employ either "Phase Control at Combiner" (PCAC), where coherence is achieved immediately upon exit of the laser system, or "Phase Control at Target" (PCAT), where coherence is achieved at the target. PCAC is the significantly easier scheme to implement, but if there exists dephasing disturbances along the propagation path, such as atmospheric turbulence, coherence at the target is lost with a corresponding reduction of intensity. PCAT schemes include means to compensate for dephasing disturbances. One known PCAT system comprises a sensor to capture an optical return from the target and a means to adjust the phase of the output beams in response to the return to maximise the amplitude of the return. Hill climbing algorithms, such as stochastic parallel gradient descent (SPGD) (see for example, WO2023 / 194921 and 7.1 KW Coherent Beam Combining System Based on a Seven-Channel Fibre Amplifier Array, Pengfei Ma et al, Optics and Laser Technology 140, 2021, 107016), are used to continuously adjust the phase of the output beams to compensate for phase changes caused by atmospheric turbulence. This is effective as long as phase adjustments to the output beams are made much faster than the time for the dephasing disturbances to cause decorrelation of the beam. These systems may need to propagate and receive feedback from light that has travelled over several kilometres. At these distances the travel time of the light is in the order of 1 O’s of microseconds, corresponding to a frequency in order of 1 O’s of kHz, whereas phase changes to the optical beams can be made at frequencies in the MHz or GHz. This means that the limiting factor in the system is the time taken to receive feedback following a phase change. As such, it is preferable to reduce the number of optimisation steps that require a feedback from die target. For hill climbing methods, the number of optimisation steps is high, usually in die 100’s to 1000’s, and as such optimisation is relatively slow. Where optimisation is performed against a continuously changing set of conditions, e.g. to optimise for atmospheric turbulence, a slower optimisation process results in reduced maximum coherence and / or mean coherence at the target. Improved optimisation speed can be achieved using machine learning mediods that function by learning a unique relationship between a light measurement and phase relationships between the beams. Known machine learning methods are trained to learn a transformation between near-to-far field images of the propagated laser array to known phase relationships between the laser beams. In doing so, for any given near-to-far intensity image, the model is able to output a phase relationship of the emitters. Note that near-to-far field images are required as they have a unique relationship with the relative optical phases between the beams. In contrast, there are multiple phase conditions that produce the same far-field light pattern, and all phase conditions produce the same near field image. Laser systems employing these machine learning methods require high resolution image information which necessitates having a feedback device, namely a camera, at the target, picking up light distribution in the near-to-far field which is achieved by placing a lens between the laser system and the camera. The camera, positioned closer to the lens than its focal length, translates the light from near field to near-to-far field. Uris arrangement is possible to implement in a laboratory environment, but in real world applications the target will often be too far away from the lasers, e.g. many km’s, to allow for high quality resolution images to be captured and / or, it would not feasible to place a detector at the target. Additionally, it is impracticable to place a lens between the laser and target in most real world applications. Further, where the beams travel through a dynamic turbulent environment to reach the target, the relationship between the phase applied by the phase controllers and the resulting intensity near-to-far field images, changes over time as the effect the turbulence has on the propagated light decorrelates. This requires retraining of the model. Atmospheric turbulence, which typically decorrelates in approximately 1ms, is orders of magnitude faster than the training process, which takes minutes to hours of time. Examples of the aforementioned type system are disclosed in Deep-Leaming-Based Phase Control Method for Tiled Aperture Coherent Beam Combining Systems; Hou Tianyue etal High Power Laser Science and Engineering, vol. 7,2019; as well as Tiled Aperture Beam Combining with Reinforcement Learning, Henrik Tunnerman and Akira Shirakawa, CLEO 2020 OSA. According to a first aspect of the invention there is provided a method of coherent beam combination of optical outputs of a tiled laser array, the method including carrying out a first optimisation process during use of the laser array comprising: using a light sensor to capture an optical intensity value of a portion of a combined beam emitted from the tiled laser array that has been reflected and / or scattered from a target through a dynamic scattering medium; altering the phases of the optical outputs through different sets of phase relationships to capture a set of intensity values; using the set of intensity values as a unique identifier of a compensatory phase adjustment for the optical outputs to increase coherence of the combined beam at the target; and adjusting the phase of the outputs to effect the compensatory phase adjustment; and carrying out a second optimisation process to compensate for a change in the dynamic scattering medium since the first optimisation process, the second optimisation process comprising altering the phases of the optical outputs through the different sets of phase relationships to capture a second set of intensity values; using the second set of intensity values as a second unique identifier of a compensatory phase adjustment for the optical outputs to increase coherence of the combined beam at the target; and adjusting the phase of the outputs to effect the compensatory phase adjustment. According to a second aspect of the invention there is provided a method of coherent beam combination of optical outputs of a tiled laser array, the method including carrying out an optimisation process during use of the laser array, the method comprising: directing a combined beam from the optical outputs of the tiled laser array through a dynamic medium towards a target; capturing a set of intensity values by: using a light sensor to capture a first optical intensity value of a portion of the combined beam emitted that has been reflected and / or scattered from the target through the dynamic medium; altering the phases of the optical outputs through one or more different sets of phase relationships and, for each of the one or more different set of phase relationships, using the light sensor to capture a further intensity value of the combined beam that has been reflected and / or scattered from the target through the dynamic medium; using the set of intensity values as a unique identifier to determine a compensatory phase adjustment for the optical outputs to increase coherence of the combined beam at the target; and adjusting the phase of the outputs to effect the compensatory phase adjustment; and repeating the optimisation process to compensate for a subsequent change in the dynamic scattering medium, each repetition of the optimisation process comprising altering the phases of the optical outputs through the one or more different sets of phase relationships to capture a further set of intensity values; using the further set of intensity values as a further unique identifier to determine a further compensatory phase adjustment for the optical outputs to increase coherence of the combined beam at the target; and adjusting the phase of the outputs to effect the further compensatory phase adjustment. The following applies to either aspect of the invention. In the most common expected application of the invention the dynamic scattering medium would be, or at least include, the earth's atmosphere, though the invention could also be applicable to other dynamic scatting mediums such as submarine environments. In a simple example embodiment comprising a tiled laser array emitting two beams, the lasers are initiated and an initial intensity value of the return recorded. A known phase adjustment is made to one beam and a known different phase adjustment is made to the other beam, and a second intensity value is recorded. Further known phase adjustments are made to the two lasers in the same manner and intensity values recorded for each. In this way it is possible to relate phase adjustment to intensity values such that for a desired intensity value a phase adjustment can be predicted and / or calculated. Where the system comprises a small number of laser emitters, e.g. two or three, it is possible to calculate the phase adjustments needed by generating and solving a set of simultaneous equations that relate phase adjustment values with intensity value. With greater numbers of lasers, the number of calculations and measurements required to solve the equations become onerous and so it is more efficient to employ a machine learning algorithm to predict the solution. As such, the set of intensity values may form an input to a machine learning algorithm, e.g. of deep learning type, trained to relate said set of intensity values to a set of compensatory phase adjustment values that increase coherence of the combined beam at the target. Central to the invention is the use of the same sets of relative phase change relationships for both the first and second optimisation processes, and where machine learning is employed, for the training of the model. Because the relationship between coherence and relative phase changes between beams accounts for the overall phase change of each beam, irrespective of its underlying physical causes, e.g. the initial phase condition of each beam and any offset applied by the atmospheric turbulence, the relationship is independent of any specific atmospheric conditions. As such, the machine learning model does not need to be retrained following decorrelation caused by dynamic turbulence. Instead the model can be executed repeatedly to optimize phase adjustments against decoherence of the combined beam at the target caused by the dynamic atmosphere. Environmental phase conditions decorrelate at a rate of approximately 1kHz, whereas the optimisation process, e.g. collecting test data and executing the model with said test data, can be carried in the order of 10s of microseconds. As such, the phase offset attributed to atmosphere turbulence remains substantially stable for the duration of optimization process and for a significant period thereafter, in which the system can operated in an optimised state. As such method may comprise repeatedly carrying out an optimisation process, e.g. from time-to-time, to compensate for a change in the dynamic scattering medium, each optimisation process comprising altering the phases of the optical outputs through the different sets of phase relationships to capture a set of intensity values; using the set of intensity values as a unique identifier of a compensatory phase adjustment for the optical outputs to increase coherence of the combined beam at the target; and adjusting the phase of the outputs to effect the compensatory phase adjustment Another advantage of using this relationship is that it is not necessary to use real world data to train tine model, rather values for the training data can be generated synthetically, e.g. randomly or pseudo randomly, which significantly reduces the burden and time required to train the model. Other advantages of using same sets of relative phase change relationships for each repetition of the optimisation processes include: (i) it provides a means to train and operate a MLM to optimise coherence using a set of single amplitude values as feedback, as opposed to a set of images. This significantly reduces that required data capture and therefore significantly reduces the time required for the model to run reducing optimisation time; (ii) feedback can be based from readings of scattered and / or reflected light that has travelled many kilometres from a target; (iii) the light sensor can be substantially collocated with the lasers and beam combiner, obviating the need for either a sensor at the target or lens between the target and system; (iv) only a single element light detector is required to capture feedback; (v) it is unnecessary to know the absolute phases of beams nor the relative phase difference between beams at any time during operation including during optimisation. Typically there is provided multiple sets (hereafter a collection of sets) of phase values, each set comprising a separate phase adjustment value, 9, associated with each optical output of the titled laser array. As such, the process of altering the phases of the optical outputs through different sets of phase relationships to capture a set of intensity values may comprise altering the relative phase of each of the optical outputs by its associated phase adjustment value of a first set and capturing a first intensity value. Then, altering the phase of each optical output by its associated phase adjustment value from a second set of phase values and capturing a second intensity value, and continuing this for each set of the collection of sets. Each set of the collection, when applied, provides a unique phase relationship between the optical outputs compared with each other set of the collection. The same collection of sets are applied, and favourably m the same order, for each repetition of the optimisation process. And where a machine learning model is trained, the same collection of sets, is used for the training as well. Although the relative phase adjustment values could be a target phase value, this is not necessary, nor preferred. Rather, it is preferably a delta phase value, i.e. the magnitude and direction by which the phase of the optical output should be changed from its current value, which for each set other than the first set will be the phase resulting from the change made by applying the previous set. This means that the actual phase value of each optical output need not be known, simplifying implementation of the system. Where the system comprises N optical beams that are combined, each set may thus comprise N phase values. However, tire phase of one of the optical outputs may be unchanged, i.e. the delta phase value remains at zero in each set, as such each set may only comprise N-l values. Preferably the tiled laser array comprises a bank of phase adjusters to allow for independent phase adjustment of each optical output prior to combination by a tiled array beam combiner, and altering the phases of the optical outputs through different sets of phase relationships is performed using the bank of phase adjusters. The optimisation process adjusts the relative phases of the beams through enough sets to provide a broad data variance through the possible phase space to provide sufficient variance in the light return from the target such that the MLM has a greater than 50% chance of generating an improved phase state within the output range of the model. A preferred way of generating tire collection is to generate the phase shift values for each set synthetically, e.g. randomly or pseudo randomly, e.g. within the limits of -n and n, using a computer implemented random number generator. By having a sufficient number of set, e.g. at least 50, a broad enough variance can be ensured. In principle, the minimum number of sets required equals the number of beams minus one, through to achieve this the values within each would need to be carefully selected to ensure the broad variance across the phase space. When the beams are initially directed at the target, it is likely the optimisation process will need to be repeated several times in immediate succession to achieve a desired degree of coherence. Determining that the beams are adequately coherent may comprise comparing the magnitude of the compensatory phase adjustments with a predetermined threshold value. In effect, this operates on the principle that corrective phase changes that are very small is indicative that the beams are substantially coherent at the target. Once optimised, and as required in claim 1, the optimisation process is repeated from time-to-time thereafter to compensate for the dynamic effects of the transmission medium of the beams or other changes in the system. For many applications, it is favourable to repeat the optimisation process at a frequency greater than 1kHz. more favourably greater than 2kHz in order that optimisation is repeated more frequently than the decorrelation frequency of the atmosphere. In the maintenance phase, because the system is already close to optimisation, any phase change adjustments required to maintain coherence at the target is likely to be relatively small compared with the initial process of reaching optimisation. In this situation, there is an advantage in using sets of phase shift values that are selected within a narrower range than used for the initial optimisation as it will significantly reduce the extent to which the combined beam is adjusted away from its optimised state when applying the phase changes required by the sets of phase relationships during the optimising process. As such, after cariying out one or more optimising processes using a first MLM trained on a first collection of sets of phase values in which the phase values of each set have values lying between a first range of values, a further optimisation process may be carried out using a second MLM trained on a second collection of sets of phase values in which the phase values of each set have values lying between a second range of values, where the first range of values is wider than the second range of values. For example, where the first range is between -re and re, the second range will be narrower than this, e.g. between -re / 10 and re / 10. According to another aspect of the invention there is provided a method of training a machine learning algorithm for optimising coherent beam combination of optical outputs of a tiled laser array passing through a transmission medium to a target, the method comprising producing training data by: constructing P sets of characteristic phase values <|> representing a transmission medium, each set comprising a number N of phase values [(^ 1 - <|>N] equal to the number of optical outputs of the tiled laser array; generating M sets of phase delta values 0 [e.g. using a random or pseudo-random number generator) corresponding to a relative phase of the emitters: each set comprising N phase delta values [91- 0N]; using each of the P sets of characteristic phase values with all M sets of phase delta values to evaluate Ep,m d” COS^0p2 ^771,2) d" *** d” Where-. P = {U.....P} m = {1,2, ...M} to produce P training sets; where Ep m represents the intensity value of light returned from the target, and At represents the amplitude contribution from each optical emitter, assumed to be substantially equal ( Ai,2,.,n=1); and training a neural network to relate each of the pth training sets as input against the pth set of characteristic phase values as a target. According to a further aspect of the invention there is provided a method of training a machine learning algorithm for optimising coherent beam combination of a tiled laser array, the method comprising (i) producing training data by: construct P sets of characteristic phase values 4> representing a transmission medium, each set comprising a number N of phase values [<|) 1 - (|)N] equal to the number of optical outputs of the tiled laser array; generating M sets of phase delta values 0 corresponding to a relative phase of the emitters: each set comprising N phase delta values [01- ON]; using each of the P sets of characteristic phase values with all M sets of phase delta values to evaluate Ep.m = (A cos(0p4 - em l) + AJ / 2 + (X2 cos(0p 2 - 0m2) + X2) / 2 + ••• + (A^ cos^pjy — 0m^ + An) / 1 Where-. p = {1,2,..., P} m = {1,2, ...M} to produce P training sets; where Ep,(i,2..m) represents the intensity values of light returned from the target for the pth training set, and At represents the amplitude contribution from each emitter, assumed to be substantially equal (Ai,2,..,n=1); and (ii) training the machine learning algorithm to relate each of the pth training sets as input against the pth set of characteristic phase values as a target. Rather than using the cosine term, a sine term could be used with the addition of a n offset term. E.g. Ar cos^<pp l — 0m>1) could be written as At sin^p^ — ± u). The P sets of characteristic phase values <|> may be generated synthetically. For example, constructing P sets of characteristic phase values <|> may comprise using a random number generator to generate the characteristic phase values ¢. The less preferred alternative is to generate a complex computer simulation of the transmission medium or alternatively, physical measurement of example transmission mediums, both far more time consuming and complex to undertake. The invention will now be described by way of example with reference to the following Figures in which: Figure 1 is a schematic of a laser system implementing a system for coherent beam combination at a target; and Figure 2 is a schematic illustrating a process of generating training data and training the model with said training data. Figure 1 illustrates a laser system 1. The laser system I comprises multiple lasers Ll-LN , a set of phase modulators 2, a tiled array beam combiner 3, an optical sensor 4, and control and processing system 5. Each laser Ll-LN outputs a coherent beam Bl-BN. The lasers Ll-LN are configured to produce beams that are spectrally matched, i.e. have substantially the same line width and central wavelength, and are power matched, namely the output powers are substantially the same. A separate phase modulator 2 is associated with each beam Bl- BN. Each phase modulator 2 is configured to control the piston phase of its respective beam Bl-BN in response to a control signal received from the control and processing system 5. The phase adjusted beams Bladj-BNadj are packed by the tiled array beam combiner 3 to provide a tiled output beam BT for travel through a turbulent medium 10, e.g. the earth’s atmosphere, towards a target 11. The optical sensor 4, implemented using a single photo detector, is configured to receive a portion of the beam BTS scattered and / or reflected from the target 11 back through the turbulent medium 10, and output an electrical signal indicative of the amplitude of the return to the control and processing system 5. The control and processing system 5 comprises a controller 51 implementing the functions of a signal controller 51A and mode selector 5 IB; a phase controller 52; and a processing system 55 comprising a store 56, a processor 57 for executing a machine learning model. The store 56 holds a set or collection 60 of M phase adjustment maps. Each phase adjustment map 60i-m comprises a set of relative phase adjustment values 50, one value for each of beams Bl-BN. Upon initiation of the laser system 1, and then from time-to-time thereafter, the mode selector 5IB switches the control and processing system 5 into an optimisation mode in which through a feedback process, the controller and processing system 5 determines optimised phase shifts to be applied to each beam B 1-BN and controls the set of phase modulators 2 to effect these phase shifts to cause the individual beams of the tiled output beam BT to combine coherently at the target 11. During optimisation, the controller 50, through the signal controller 51A passes a first phase adjustment map to the phase controller 52 which in response operates the phase modulators 2 to alter the relative phase of the beams B 1-BN by the degree specified by the first phase adjustment map. The absolute relative phases of the beams Bl-BN do not need to be known. For example, if the absolute phase value of B1-B3 happened to be (whether it was known or not) k, n / 4, 2^ / 3, a phase adjustment map comprising the relative phase change values a / 2, it, -ti / 4, when applied, would transform the phase of the beams to: 7i / 2, -3ti / 4, 1 Itt / 12. The amplitude of the return from the target 11 sensed by the optical sensor 4 is stored in an array 57 of length M in store 56 in which each element of the array is associated with a different phase adjustment map of the collection 60. This is repeated for each phase map of the collection 60 (to produce M amplitude values) such that a separate amplitude value is associated with each of the M phase adjustment maps of tire collection 60. In each instance the phase modulators 2 may alter the relative phase of the beams B1 -BN by the degree specified by the phase adjustment map from its setting resulting from the transformation made by applying the previous phase map of the collection 60. Where the target is several kilometres from the laser system 1, the relative phases of the beams B1-BN can be switched far more quickly than it takes for the phase adjusted beams Bladj-BNadj to travel to the target and reflect back to the optical sensor 4. As such, in practice, the system 1 does not wait to receive a return before applying the next phase adjustment map to the phase modulators 2, but rather cycles through each of phase maps and populates the array 57 with amplitude values in the order they are received, which naturally aligns with the order the phase adjustment maps were applied. This significantly reduces the time required to collect the amplitude values. Following application of each phase adjustment map 60, the controller 51 causes the phase controller 52 to revert to the phase values prior to applying the training phase adjustment maps. This can be achieved by subtracting the sum of the phase adjustments made to each beam through applying all the phase maps 60. Following collection of the amplitude values, the processor 57 executes the machine learning model (MLM) which takes each phase map of the set 60 with corresponding amplitude values from the store 60 as input, and from them calculates an optimising phase adjustment map 61 as an output. The optimising phase adjustment map 61 comprises a set of relative phase values, one for each beam Bl-BN, that expresses the direction and magnitude of phase change to be imparted on each beam to increase coherence of the tiled output beam BT at the target 11. To determine if the output beam BT is sufficiently coherent at the target 11, the optimising phase map 61 is compared with a threshold. For example, the magnitude of each phase value of the optimising phase map 61 may be compared with a threshold value, or the sum of the phase values may be compared with a sum threshold value. If the values(s) fall below the threshold, indicative that the size of the phase adjustments made by applying the optimising phase map 61 are small, the system will be considered optimised, i.e. the output BT is coherent at the target 11. If the optimising phase map 61 is above the threshold, then the optimisation process described above is repeated. Starting with each beam Bl-BN in its phase state set through applying the optimising map 61, the phase modulators 2 alter the relative phase of the beams Bl-BN by the degree specified by each phase adjustment map of the collection 60, in the same order to collect a new set of amplitude values comprising a separate amplitude value is associated with each phase adjustment map of the set M. Again, following application of each phase adjustment map 60, the controller 51 causes the phase controller 52 to revert to the phase values prior to applying the phase adjustment maps 60. Following collection of the new set of amplitude values, the processor 57 re-executes the machine learning model (MLM) which takes each phase map of the set 60 with the corresponding new amplitude values from the store 60 as input and from them calculates a revised optimising phase adjustment map 61 which is applied to the beams Bl-BN. The values of the revised optimising phase map 61 are compared with the threshold, and the optimisation process repeated until a revised optimising phase map is generated that falls below the threshold. Once it is determined that the output BT is coherent, the system 1 may switch to a maintenance routine in which the optimisation process is repeated from time-to time to maintain coherence of the output beam BT at the target 11. The optimisation process may be repeated periodically, where so it is favourably repeated at a frequency that is greater than decoherence frequency of the transmission medium, and thus typically greater than 1kHz. Alternatively, it may be repeated dynamically, namely in response to a reduction of the amplitude of the return signal received by the detector 4 over a threshold amount. The process of producing the training data for training the MLM mentioned above, and training of the MLM is described with reference to Fig 2. The set of M phase adjustment maps 60 used in the optimisation process described above are generated. Each phase map 61, represented by a single column of the array, comprises a set of N phase shift values 80i-n, one for each of the number of beams of the laser system 1. Hie phase values maps are generated with values to provide, taking the set of maps as a whole, a broad data variance through the possible phase space to enable the MLM to provide optimised phase maps with high confidence values irrespective of the phases of the beams Bl-BN. The phase maps must provide sufficient variance in the light return from the target such that the MLM has a greater than 50% chance of generating an improved phase state within the output range of the model. The currently preferred way of generating the phase maps is to generate the phase shift values for each phase map randomly or pseudo randomly, e.g. using a computer implemented random number generator, within the limits of -n and z. By having a sufficient number of maps, e.g. M=50, a broad enough variance can be ensured. Alternatively, it would be possible to select combinations of phase difference values for each set in order to minimise the total number of sets required to provide an acceptable confidence value. In principle the minimum number of phase maps required equals N-l, though the values for 80 within each map 61 would need to be carefully selected to ensure the broad variance across the phase space. A set 70 of target phase adjustment maps 71 are also synthetically generated. Each target phase adjustment map 71, represented by a single row of tire array comprises a set of N phase shift values 84>i-n, one for each of the number of beams Bl-BN of the laser system 1. The phase shift values for each target phase map 71 are generated randomly or pseudo randomly, such as by a computer implemented random number generator, within the limits of -rt / 4 and n / 4. This range provides a good balance between optimisation speed and accuracy. A wider range reduces the confidence of the deep learning network and a smaller range improves the confidence but requires more iterations to optimise. The number P of target phase maps 71 is selected to provide enough training data to provide an acceptable level of model accuracy, as will be familiar to those skilled in the art of machine learning techniques. Again, the phase shift values for the target phase maps 71 need not be generated randomly but it may be that more target phase maps 71 would be needed to provide sufficient variance. The equation below relates the amplitude of a combined beam with both the relative phase of the component beams, and the phase shifting effects caused by the medium through which the beams propagate to the target. E1 = Ar COS(0! — 0x) + A2 cos(02 ~ #2) d-----b COS(<pN — 9N) A represents the intensity value of light returned from the target (and thus representative of coherence at the target), and the cosine terms Ar cos(0x — to An cos(0w — 0N) represent the intensity value from each of the respective beams BIBN. When the phase 0 of a beam is conjugate to the phase shift cp applied by the medium to that beam at its arrival at the target 11, the cosine term associated to that beam is maximised. A maximum value of E is realised by maximising tine value of each term, corresponding to perfect coherent beam combination at the target 11. The above equation is evaluated to determine a first value for E using a first target phase adjustment map 711 from the set of size P target phase adjustment maps 70 and the first phase adjustment map 61i from the set of size M phase adjustment maps 60. This is repeated for each remaining phase adjustment maps 61 of the set 60 to generate a set of size M values for E. Hie set of M values for E (E1:1.M) together with the first target phase adjustment map 711 forms a single instance of training data. This process is repeated for each target phase adjustment map 71 of the set 70 to generate P instances of training data each comprising a set of M values for E, generated from the set of phase adjustment maps and the selected target phase adjustment map 71. The MLM is trained using the E values as an input and the target phase adjustment maps as the target output of the model by which a reduced error against is sought. The end result of the training process is a model that transforms any given set of E values into an optimised phase adjustment map. To avoid the possibility of negative terms for the component values and thus Ep m, a refined form of tine equation may be used as provided below: Ep.m = (^1 COS^p t — ^m,l) + ^1) / 2 + (A2 COS^0p 2 - dm,2) + ^) / 2 + ••• + dm+ 4^,) / 2 Further, rather than using the cosine, a sine term could be used instead with the addition of a ri offset term. For example, Aa cos^<pp l — 0m^ could be written as A^ sin($pl i andi42 cos^<^>p 2 ^771,2) ^2 dm,2 i Accuracy of the transform relies on the same phase adjustment maps being used to generate the M values for E, though the order may be changed so long as it is known. Advantageously this training method does not require complex computer simulation of light propagation or physical measurements of light propagation. Examples of machine learning model architectures suitable to implement the model include deep learning neural networks, one dimensional convolutional neural networks or combinations thereof. The detector 4 could comprise an array of photo detectors whose outputs are combined. Rather than using multiple lasers, the system may comprise a single seed laser that produces a seed beam which is divided to provide beams Bl-BN. Typically such a system would also include a set of amplifiers to amplify each of beams Bl-BN. As previous mentioned, once optimised, the system will repeat the optimisation process, e.g. from time to time. Repetition may be configured to occur periodically or dynamically. A potential problem of optimising periodically is that the optimisation process may be carried out either more often than required or, less likely, not frequently enough. A problem with doing it more often than required is that each time the optimisation process is carried out it is necessary to adjust the phase of the beams by the phase maps 60 which will make a relatively optimised output beam less coherent for the duration that this is done. Carrying out the process dynamically e.g. in response to a reduction of the amplitude of the reflected beam at the detector 4 is that this necessitates reduction in coherence to be sensed. A possible means to address this problem is to use two machine learning models; the model previously described to bring the system into a initial optimised state, and then switch to use a second model to maintain the optimised state with changes in the dynamic transmission medium. The second model is substantially the same except it is trained on a second collection of sets phase adjustment maps that differs from the first model in that the phase shift values for each phase map are selected (e.g. using the same technique used to generate the phase shift values of the first model) within limits that are narrower than the first model. Thus in the example where the first model uses a range of -n and n, the second model will be narrower than this e.g. -7i / 10 and 7t / 10. A narrower breadth will cause less deviation of the output beam from the previously optimised state when collecting the intensity values for inputting to the second model when running the optimisation process.

Claims

1. A method of coherent beam combination of optical outputs of a tiled laser array,the method including carrying out an optimisation process during use of the laser array, the method comprising:directing a combined beam from the optical outputs of the tiled laser array through a dynamic medium towards a target;capturing a set of intensity values by:using a light sensor to capture a first optical intensity value of a portion of the combined beam emitted that has been reflected and / or scattered from the target through the dynamic medium;altering the phases of the optical outputs through one or more different sets of phase relationships and, for each of the one or more different set of phase relationships, using the light sensor to capture a further intensity value of the combined beam that has been reflected and / or scattered from the target through the dynamic medium;using the set of intensity values as a unique identifier to determine a compensatory phase adjustment for the optical outputs to increase coherence of the combined beam at the target; andadjusting the phase of the outputs to effect the compensatory phase adjustment; andrepeating the optimisation process to compensate for a subsequent change in the dynamic scattering medium, each repetition of the optimisation process comprisingaltering the phases of the optical outputs through the one or more different sets of phase relationships to capture a further set of intensity values; using the further set of intensity values as a further unique identifier to determine a further compensatory phase adjustment for the optical outputs to increase coherence of the combined beam at the target; andadjusting the phase of the outputs to effect the further compensatory phase adjustment.

2. A method according to claim 1 wherein the set of intensity values forms an input to a machine learning algorithm trained to relate said set of intensity values to a set of compensatory phase adjustment values that increase coherence of the combined beam at the target.

3. A method according to claim 1 or 2 wherein the different sets of phase relationships are pre-determined.

4. A method of training a machine learning algorithm for optimising coherent beam combination of a tiled laser array, the method comprising(i) producing training data by:construct P sets of characteristic phase values <|) representing a transmission medium, each set comprising a number N of phase values [¢1- <|)N] equal to the number of optical outputs of the tiled laser array;generating M sets of phase delta values 0 corresponding to a relative phase of the emitters: each set comprising N phase delta values [01- 0N];using each of the P sets of characteristic phase values with all M sets of phase delta values to evaluateEp^ 4" C®s(jPp,2 ^m,2^) 4" *** 4" ^^^(jPp,N ^m,N^Where:p = {1,2,...,P]m = {1,2, ...M]to produce P training sets; whereEp,represents the intensity values of light returned from the target for the pth training set, and At represents the amplitude contribution from each emitter, assumed to be substantially equal ( Ai,2,..,n=1); and(ii) training the machine learning algorithm to relate each of the pth training sets as input against the pth set of characteristic phase values as a target.

5. A method of training a machine learning algorithm for optimising coherent beam combination of a tiled laser array, the method comprising(i) producing training data by:construct P sets of characteristic phase values <|) representing a transmission medium, each set comprising a number N of phase values [¢1- <|)N] equal to the number of optical outputs of the tiled laser array;generating M sets of phase delta values 0 corresponding to a relative phase of the emitters: each set comprising N phase delta values [01- 0N];using each of the P sets of characteristic phase values with all M sets of phase delta values to evaluateEp,m = (A cos(0p4 - em l) + AJ / 2 + (X2 cos(0p 2 - 0m2) + X2) / 2 + ••• + (An cos^pjy — 0mjv) + ^n) / 2Where-p = {1,2,..., P]m = {1,2, ...M}to produce P training sets; whereEp,(i,2..m) represents the intensity values of light returned from the target for the pth training set, and At represents the amplitude contribution from each emitter, assumed to be substantially equal (Ai,2,..,n=1); and10 (ii) training the machine learning algorithm to relate each of the pth training sets as input against the pth set of characteristic phase values as a target.

6. A method according to claim 4 or 5 wherein constructing P sets of characteristic phase values o comprises using a random number generator to generate the 15 characteristic phase values <|).

7. A laser system comprising:means for generating multiple coherent beams;a tiled array beam combiner to combine the multiple coherent beams to provide a combined beam;a light sensor configured to sense portion of a combined beam emitted from the coherent beam combiner that has been reflected and / or scattered from a target through a dynamic scattering medium;multiple phase modulators, a separate phase modulator of the multiple phase modulators associated with each of the multiple beams;an optimisation and control system comprising:a store comprising a set of phase maps, each phase map comprising a set of relative phase adjustment values, one value for each beam;and in which the optimisation and control system is configured tocany out a first optimisation process during use of the laser array comprising:capturing a set of intensity values by:using the light sensor to capture a first optical intensity value of a portion of a combined beam emitted from the tiled laser array that has been reflected and / or scattered from the target through the dynamic scattering medium;altering the phases of the multiple coherent beams through one or more different sets of phase relationships using the set of phase maps and, for each of the one or more different set of phase relationships, using the light sensor to capture a further intensity value of the combined beam that has been reflected and / or scattered from the target;execute a predictive model trained to use the set of intensity values as a unique identifier to determine a compensatory phase adjustment for the optical outputs to increase coherence of the combined beam at the target; andcontrolling the phase modulators to adjust the phase of one or more of the multiple beams to effect the compensatory phase adjustment; andrepeating the optimisation process to compensate for a change in the dynamic scattering medium, each repetition of the optimisation process comprising5 altering the phases of the multiple coherent beams through die one or more different sets of phase relationships using the set of phase relationship maps to capture a further set of intensity values;executing the predictive model using the second set of intensity values as a further unique identifier to determine a further compensatory phase adjustment for the optical10 outputs to increase coherence of the combined beam at the target; andcontrolling the phase modulators to adjust the phase of one or more of the multiple beams to effect the further compensatory phase adjustment.28

Images