Using Deep Machine Learning to Understand the Physical Performance Bottlenecks in Novel Thin‐Film Solar Cells

There is currently a worldwide effort to develop materials for solar energy harvesting which are efficient and cost effective, and do not emit significant levels of CO2 during manufacture. When a researcher fabricates a novel device from a novel material system, it often takes many weeks of experimental effort and data analysis to understand why any given device/material combination produces an efficient or poorly optimized cell. It therefore takes the community tens of years to transform a promising material system to a fully optimized cell ready for production (perovskites are a contemporary example). Herein, developed is a new and rapid approach to understanding device/material performance, which uses a combination of machine learning, device modeling, and experiment. Providing a set of electrical device parameters (charge carrier mobilities, recombination rates, trap densities, etc.) in a matter of seconds thus offers a fast way to directly link fabrication conditions to device/material performance, pointing a way to further and more rapid optimization of light harvesting devices. The method is demonstrated by using it to understand annealing temperature and surfactant choice and in terms of charge carrier dynamics in organic solar cells made from the P3HT:PCBM, PBTZT‐stat‐BDTT‐8:PCBM, and PTB7:PCBM material systems.

turn it into electricity. [1] Solar cells based on silicon currently dominate the market, however these cells require silicon of 99.9999% purity or more. Obtaining this grade of silicon from naturally occurring silicon dioxide, is a complex, expensive, and multistep process, involving liquefaction of silicon, heating to temperatures in excess 1000 K, and numerous gasification and purification steps. [2] This means, that significant amounts of energy have to be invested [2,3] before any financial returns are seen from a silicon based photovoltage (PV) system, furthermore it can often take years before a PV system can recoup the greenhouse gases emitted during its manufacture. [4] For these reasons, the last two decades have seen an intense search for novel nonsilicon based light harvesting materials. [5] Some of the more promising materials developed as alternatives to silicon include; semiconductors based on conducting organic molecules, [6,7] copper indium gallium selenide (CIGS), [8,9] and more recently perovskite [10,11] based materials, which captured the imagination of the community. Although, none of these material systems have yet become a serious competitor to silicon, [12,13] there remains an intensive and on-going effort to replace silicon, with a potentially as of yet undiscovered semiconductor. [14] Early in the development of a new material system it is not known how to optimize it for use in solar cells. This means in dozens of labs around the world, device engineers, and scientists, build hundreds of thousands of candidate devices, trying to optimize the material system and boost the solar energy conversion efficiency. Translating a promising solar cell material to an optimized device can take upward of 10 years of worldwide scientific effort. Key to this development process is the need to quantify and understand the physical mechanisms behind the solar cells operation [15] which in turn dictate its efficiency. Specifically, charge carrier recombination, [16] trapping, [17] charge carrier mobility, [18] and charge generation [19] all have to be measured throughout the cells life time and understood. However, measuring these quantities is nontrivial, because often material parameters change considerably as a function of film thickness, device structure, and deposition conditions. [20] Furthermore, many novel materials have high numbers of trap states thus changing the applied voltage bias or light intensity leads to significant changes in mobility and recombination

Introduction
Over the last few decades, there has been considerable academic and industrial attention focused on developing low cost and low carbon technologies to harvest the suns energy and rates [21] making it hard to benchmark one cell/device structure/ material system against another. In the more extreme case of a perovskite devices, a sea of ions within the device slowly moves as a voltage is applied fundamentally altering the device's behavior, [22] and making it yet harder to study and improve its operation.
Consequently, as the field of thin film solar technology has been developed, so has an array of measurement techniques to probe the fundamental material properties of these materials. Examples are dark current density-voltage (JV) measurements which have been used to measure energetic disorder; [23] charge extraction by linearly increasing voltage (CELIV), which has been used to measure effective mobility; [24] time of flight (ToF) photocurrent measurements for mobility and energetic disorder determination; [25][26][27] transient photovoltage (TPV) for charge carrier lifetime measurements; [28] transient photocurrent (TPC) again for effective mobility and energetic disorder; [29] space-charge-limited current (SCLC) for electron and hole mobility; [30] and impedance spectroscopy has been used to measure recombination, mobility, and contact quality. [31] In general, to extract a physical parameter from any of these measurements, one needs to fit an analytical model to the experimental data. One simple example of this approach is to use the slope of dark JV curve just after diode turn on to extract the diode ideality factor (n), which in turn can be used as a measure for the disorder within a device. One usually fits the Shockley equation to the JV curve to extract this quantity. Where n is the ideality factor, V is the applied voltage, J is the diode current, J 0 is the reverse saturation current density, and all other parameters take their usual meanings. However, recently Würfel et al. demonstrated that the ideality factor could also be strongly influenced by mobility, [32] making it difficult to fully trust the application of Equation (1). Another example would be the CELIV technique for determining charge carrier mobility. This was first demonstrated by Juška in 2000, [24] Deibel et al. then revised the method with an updated equation, [33] and since then there have been other papers published either suggesting improvements on the method or questioning its validity in certain regimes. [34] In general, the drawbacks of this general approach of fitting analytical models to experimental data are fourfold; 1) The most elucidating measurements (i.e., ToF) can be time consuming and complex to perform; 2) analysis of the results using analytical models can be time consuming; 3) it is not always clear if the analytical models are valid for a new material system and under what experimental conditions they are valid (e.g., carrier densities, carrier gradients, light intensities, etc.); and 4) often nonstandard devices have to be fabricated to perform the measurements (e.g., ToF), thus there is uncertainty as to how the results relate back to a working solar cells. Another approach to extract physical material parameters from experimental data is to fit a Monte-Carlo or finite difference based drift-diffusion device model to the data. The advantage of this approach over using analytical models, is that it can bring significant insights into device operation: one can for example examine carrier profiles across the device as a function of time or voltage. However to use these models again requires a significant investment of both human and computational time, typically fitting such a model self consistently to experimental data requires a week of computational time on a small cluster, and a few days of time from an expert in device modeling to setup and understand the simulation results (see the Supporting Information for a more detailed discussion of this topic). This makes these methods unsuitable for a high throughput lab environment.
The factors outlined above often prevent scientists (especially in smaller labs) from carrying out detailed analysis of their freshly fabricated devices and understanding which material parameters (mobility, recombination rates, trap densities, etc.) are limiting device performance. Progress in the field is therefore slowed as scientists rarely fully understand why a device functions as it does.
In this work, we propose a new approach to link solar cell performance to microscopic material parameters. Rather than attempting to analyze experimental data directly using analytical or numerical models, we demonstrate a method based on machine learning and deep artificial neural networks. We first use a numerical device model to generate simulated JV, and transient current/voltage curves from nominally good and nominally bad solar cells. We then use these curves to train deep artificial neural networks to identify the physical reasons for performance bottlenecks in any given device and to extract material parameters directly from the simulated data sets. We find the deep neural network is able to calculate material parameters such as mobility, number of trap states, recombination time constants, and parasitic resistances from standard light and dark JV curves within a few seconds, without the need for complex time domain measurements or analysis. We then use this new method to identify performance bottlenecks in devices made from P3HT:PCBM, PTB7:PCBM, and the novel PBTZT-stat-BDTT-8:PCBM material system. The result is a general and efficient turnkey method for identifying factors limiting device performance in a lab setting, this novel method will enable scientists to more quickly optimize novel material systems and thus speed up the search for a replacement for silicon solar cells.

Experimental Section
Organic photovoltaic (OPV) devices were fabricated on prepatterned indium tin oxide (ITO)-glass substrates, the substrates were cleaned using acetone, isopropanol, and deionized-water in an ultrasonic bath. A layer of commercially available aluminium zinc oxide (AlZnO, Nanograde) was then applied as a uniform coating by doctor blade at 40 °C. The AlZnO was then annealed at 100 °C for 10 min in air. The active layer material PBTZT-stat-BDTT-8:PCBM was prepared by fully dissolving it in a solution of dichlorobenzene at 30 mg cm −3 , this solution was then blade-coated in air to achieve a thicknesses of either of 150 or 350 nm, thickness values were obtained using a profilometer. A 2 min drying period at 60 °C on a hot plate followed to ensure removal of any residual solvent. Next 0.1 mL of PEDOT:PSS (Heraeus HTL 4083) was spread and uniformly coated by doctor blade at 70 °C. Finally, Ag (100 nm) cathodes were thermally evaporated through a shadow mask to define the cell structure. The chemical structure of PBTZT-stat-BDTT-8 is depicted in Figure 1a, while the energy diagram of the device is shown in Figure 1b, and the physical structure of the device is depicted in Figure 1c.
Current-voltage characteristics were measured using a Keithley 2400 SMU both in the dark, and while cells were illuminated by a Newport Solar Simulator at 100 mW cm −2 . The corresponding JV curves are plotted as dots in Figure 2. All characterization was performed in a dry-nitrogen atmosphere.

Method
Over the last 20 years machine learning (ML) approaches have been increasingly used to accelerate material science discovery, [35][36][37][38][39][40][41][42][43][44] applications include screening crystal structure, [45] rapid searching for thermometric materials, [46,47] predicting material properties from their structure, [48] predicting crystal structure, [36] and screening polymers for energy harvesting applications. [49,50] Until recently however neural networks that mimic the learning process of biological neurons have been a relatively unsuccessful class of machine learning algorithm, as they underperformed most other techniques for machine learning and data classification. [51,52] This changed recently with two important technological developments. [53] First, the increased availability of massive data sets on which the neural networks can be trained. These data sets have become available due to the exponential growth in labeled (meaning categorized by a human) images and audio data. For example, if one searches for the phrase teapot in an online shopping portal, one will obtain hundreds of images of "teapots," which have all be identified as such by humans. Large data sets such as these were not available only few years ago. Second, to train large neural networks on these new large data sets, massively parallel computing platforms are needed, with the recent rise of graphics processing unit (GPU) computing, and more recently dedicated neural network chips, training large networks on large data sets has become significantly easier. [53] At the heart of the artificial neural networks is the artificial neuron; [54] this is an attempt to represent a biological neuron in mathematical notation, as shown in pictorial form in Figure 3b. The artificial neural network has a set of input values, A 1 … A n into which data is fed, and set of weights W 1 … W n by which the input values are multiplied. Finally the values A n W n are summed. Thus, the value of x in Figure 3b is given by The value x is then acted upon by an activation function f(x), the simplest of which sets the output of the neuron: to 1 if x is above a given value, and sets the output 0 if the value of x is below a given value. In this way the neuron can make simple decisions based on the input values it receives. In this work we use the rectified linear activation function which has a strong biological basis. [54] Just as in biological systems, these neurons are joined together to form networks of neurons which can be used to classify data or make complex decisions. The network we use in this work is depicted in Figure 3a (to simplify the diagram only one in ten neurons in each layer is drawn). Data in this network flows from left to right. Data extracted from device measurements is presented to the network on the left hand side at the input layer (red dots). The network works on this information using its neurons (blue dots), and the estimated values of material parameters such as electron and hole mobility (µ e/h ), number of trap states (N e/h ), or shunt resistance (R shunt ) are given on the output layer (green dots).
Before a neural network can be used to analyze device data it must be trained. Key to successfully training neural network is to have a large data set. Using experimental data for training is problematic for two reasons; a) tens of thousands of JV curves measured from different devices will be needed; and b) it is often very difficult to know what the material parameters such as mobility, recombination rates really are for any given device. For these reasons we use a Shockley-Read-Hall based drift-diffusion model (the general-purpose photovoltaic device model, gpvdm) to produce the training data. [20] We set up the device layer structure as described in the experimental section of the paper. The highest occupied molecular orbital (HOMO)/ lowest unoccupied molecular orbital (LUMO) levels, and real/ imaginary parts of the refractive index were manually entered into the model for each material layer. Using this base structure, we then generated a set of 20 000 devices, each with randomly assigned electrical parameters, including random recombination time constants, carrier trapping rates, trap Adv. Funct. Mater. 2020, 30,1907259  In a lab setting a light JV and dark JV curve would almost always be measured on a newly fabricated device however, more complex time domain measurements such as CELIV or TPC would only be performed on a few selected devices to extract extra information as these measurements are more complex and time consuming to perform. By simulating a range of Adv. Funct. Mater. 2020, 30,1907259   . a) A diagram of the neural network used to extract material parameters from the data within this paper, the actual network used had ten times more neurons in each hidden layer than the diagram depicts but otherwise the same structure. Visible on the left hand side of the image is the experimental (or simulated) data, with the red dots on the curves representing the points at which the curves were sampled to from input vectors for the neural network. b) In the diagram a light and a dark JV curve are each being sampled at 6 places to provide 12 data points to the neural networks 12 input nodes. Any number or combination of experimental measurements can be placed on the input to the network, one simply has to extend the number of input neurons, and retrain the network. The neural network itself has red input nodes, blue hidden layers, and green output nodes. Each output node corresponds to a device/material parameter such as charge carrier mobility or trap density. Inset: A single neuron. steady state and transient experimental measurements we will be able to choose which ones are used to train the neural network and understand which experiments provide most accurate device/material parameters.
Each JV curve or transient from a simulated measurement is effectively two columns of numbers containing one hundred to a few thousand data points. To cut down the amount of data the neural network has to process, we turn each measurement into a series of 6-12 data points. For the JV curves, this means sampling the current at 0.0, 0.1, 0.2,…, 0.7 V and storing the values in a vector before it is fed into the neural network. The process of sampling the JV curves, and feeding the values into the neural network is shown on the left hand side of Figure 3, where both a light and dark curve are being fed into the network. In a similar way any combination of time domain or steady state measurements can be fed into the neural network, one simply has to adjust the number of input nodes accordingly. We chose the number of data points to represent each measurement by balancing the need to have a small data set which is fast to train, and the need to have enough points to describe each curve accurately. Finally, we should point out that for this process to be effective, the device model must be a fair representation of a real device, the device model we use has previously been validated against multiple experimental data sets and proven to be predictive. [20,[55][56][57] More details on the implementation can be found in the Supporting Information.
The key advantage to using this neural network/machine learning approach is that once the neural network is trained, it can be applied to an unlimited number of similar devices with minimal computational overhead, whereas with approaches base on fitting physics based device models to experimental data, significant computational resource is required each time a new device is examined. We are in effect paying the computational cost once up front rather than repeatedly at the time of use (this is discussed at length in the Supporting Information). A further advantage of our approach is that it separates the task of running device models/training neural networks from their point of use in the lab environment. One could therefore envisage an open online database where scientists could exchange their pretrained networks so they could be reused. Table 1 displays how accurately the neural network can extract a given material parameter from a set of experimental data. The first column on the left displays the parameter extracted from the material system. Both electrons and holes are present in the devices, thus there will be two possible values of free carrier mobilities which could be extracted from the data. However, unless special electron or hole only devices are fabricated and measured, it will not be possible to assign a measured mobility to one carrier species. Therefore the neural network outputs three values for mobility, a maximum value (µmax free ), a minimum value (µmin free ), and an average value µ avg free . The maximum and minimum values can be assigned to either the electron or hole mobilities, but from our approach we cannot determine which one, unless we have a priori knowledge, as to which material is likely to be more mobile. Often in polymer:fullerene blends one can assume the electron transporting fullerene will be more conductive than the hole transporting polymer. Both Adv. Funct. Mater. 2020, 30,1907259  The value of mobility can vary by up to five orders of magnitude in an organic semiconductor device, therefore to describe the networks ability to extract mobility from any given data set, we give its accuracy using a log scale. Thus a value of 1.0 in the top three rows of the table would means the mobility can be extracted from the experimental data within one order of magnitude, where as smaller values mean more accurate extraction. A similar approach is taken for values of trap density which can vary by up to 20 orders of magnitude in a real device and shunt resistance which can vary by up to six orders of magnitude in a device. Contact resistance and tail slope are not calculated on a log scale.

Results
By looking at Table 1, it can be seen that the network extracts average and maximum free charge carrier mobility (µ avg free , µ max free ) within less than half an order of magnitude from a dark/light JV curve, and over one order of magnitude from a CELIV/TPV transients. Of the measurement techniques investigated transient photocurrent is able to extract the minimum free charge carrier mobility (µ min free ) most accurately to within half an order of magnitude. When JV dark , JV light , TPC −1 V , TPV, and Suns-V oc measurements are combined as inputs to the network the values of mobility can all be extracted to within half an order of magnitude.
It might be somewhat surprising that the network finds it difficult to extract mobility from a CELIV transient, however there has been some discussion in the literature suggesting CELIV is a difficult way to measure mobility. [34] Values of tail slope (E min , E max ) can be extracted from most measurements within around 14 meV, while values of trap density can be extracted to within about 1.5 orders of magnitude. Unsurprisingly, the network can extract the value of shunt resistance almost exactly from a dark JV curve. The final two parameters at the bottom of the table, µ avg max P , P avg max τ , represent the average mobility, and average recombination time constant at the maximum power point under AM1.5 solar radiation. It is interesting to note that the network can extract the value of P avg max τ quite reliably even from the dark JV curve alone. Figure 5 visualizes the results from the JV dark , JV light , TPC −1 V , TPV, Suns-V oc column of Table 1. The horizontal axis represents the true value of a given device parameter, while the vertical axis represents the neural network's prediction of the parameter, all 20 000 devices in the data set are plotted on the graphs. Ideally the plots should be a perfect diagonal, any off diagonal elements represent nonperfect estimation of device parameters. If one examines the curve for R shunt , it can be seen that it is a near perfect diagonal meaning extraction of this parameter is very good. The plots for mobility (µ avg,max,min ), recombination constant (τ avg ) also lie close to the diagonal. More spread out but clustered around the diagonal are N avg and N max . The tail slopes seem the hardest to extract from the simulated data, E max,avg appear to be broadly clustered around the x/y axis, while E min , forms a triangle in the upper left of the plot. Again extracting the most shallow tail slope would be expected to be most physically difficult as any current from Table 1. A summary of the neural network's ability to extract material parameters from dark JV curves, light JV curves, CELIV transients, TPC transients, TPV transients, and a combination of the aforementioned data sets. The numbers given are the average errors the network produces when estimating a material parameter. Material parameters which can vary by multiple orders of magnitude are given on a log scale (see the far right hand column of the table for units). Before learning commences it is usual to initiate a neural network's weights with random numbers, consequently there is some inherit noise in the learning process and the final quality of the neural network. Therefore, we split the data set into five parts and in turn train the network four of the five sets while testing on the remaining part. This process is performed five times, and the average presented in the table above, this is process is formally referred to as fivefold cross validation. This data corresponds to the thick device. These are RMS errors to the figure caption. True versus predicted curves can be found in the Supporting Information, along with detailed discussion.  it will be masked by the current from the trap states with the broader tail. [58] The trained neural network was then applied to the experimental data from both the thick and thin PBTZT-stat-BDTT-8:PC 61 BM devices shown in Figure 2. The extracted material parameters are shown in Table 2, it is interesting to note that the parameters extracted for both devices are quite similar as would be expected as both devices are made of the same material system, except for the values of mobility and recombination constant at P max , which one would expect to be structure dependent. The minimum value for the tail slope of the trap states is 53 meV, while the maximum value is 80 meV. Although these are the first reports of energetic disorder in the PBTZT-stat-BDTT-8:PC 61 BM material system, they compare well to previously reported values for other similar material systems. For example, the tail slopes in P3HT:PC 61 BM have previously been reported as 35-40 meV for the LUMO, and 60-65 meV for the HOMO. [29,55] We therefore attribute 53 meV to the electron tail slope and 80 meV to the hole tail slope.
Currently, there is only one previous report of charge carrier mobilities in the PBTZT-stat-BDTT-8:PC 61 BM material system, [30] where SCLC was used in electron and hole only devices. Values of mobility were determined of around 1 × 10 −7 m 2 V −1 s −1 and of around 1 × 10 −8 m 2 V −1 s −1 for electrons and holes respectively. We determine the maximum and minimum values of mobility as 9 × 10 −7 and 4 × 10 −8 m 2 V −1 s −1 , while the value of effective mobility at µ max P is estimated as Adv. Funct. Mater. 2020, 30,1907259 Table 1. The horizontal axis represents the true value of a given device parameter, while the vertical axis represents the neural network's prediction of the parameter, all 20 000 devices in the data set are plotted on the graphs. Ideally the plots should be a perfect diagonal, any off diagonal elements represent nonperfect estimation of device parameters by the neural network.
3 × 10 −8 m 2 V −1 s −1 for the thick device and 1 × 10 −6 m 2 V −1 s −1 for the thin device. A higher mobility would be expected for a thin device, as the center of the device would be closer to the charge injecting contacts, and thus the traps would be expected to be more filled. The values of trap density (N) are given as 2 × 10 24 m −3 as a maximum and 7 × 10 20 m −3 as a minimum. These are comparable to the values previously given for a P3HT:PCBM material system. [29] The given value of R shunt reproduces the low voltage region of the dark JV curve and thus can be considered correct although the value is low for a highly efficient device. The values of R contact at 40 Ω is reasonable although high for an organic solar cell. The other values in the table are discussed in the next section, where we compare these results to detailed numerical modeling.

Validation of Neural Network Output
In order to understand how accurate our neural network's values are in Table 2, we fit the device model used in the first half of the paper directly to the experimental data shown in Figure 2. For both the thick and thin device we use the same parameter set, but simply adjust the thickness of the active layer. It should be noted that in comparison to the neural network method presented in this paper, fitting a device model to experimental data is extremely computationally expensive, requiring a cluster of 50 CPUs. This is because to evaluate each improved guess of the parameter set the model must be rerun. The results of this fitting procedure are shown in Table 3.
If one compares the values in Tables 2 and 3, one can see the values are close, with mobility values being given at around 1 × 10 −7 m 2 V −1 s −1 , the trap density being given to around 1 × 10 23 m −3 , the shunt resistances agree to well within an order of magnitude, the values for R contact disagree only by 8 Ω, but it is always hard from JV curves to distinguish which part of the resistance should be attributed to the contact and which to the internal resistance of the device. Both the recombination time constants ( Pmax τ ) and the mobility at P max also agree well. Finally, the fit to the device model predicts a value of 50 meV for the tail slope, which the neural network predicts a range from 53 to 80 meV.

Understanding the Influence of Thermal Annealing and Additives on Device Fabrication
After an organic solar cell is fabricated it is often thermally annealed, this allows the crystalline domains to grow or shrink and for molecules to rearrange their orientation. Thus thermal annealing represents a useful post fabrication method to optimize material microstructure for higher efficiency devices. Another method used to increase device efficiency is to introduce additives into the solution containing the active layer before it is deposited, these additives typically act as selective solvents for the fullerene molecules and thus alter the film formation process. Although the use of annealing and additives to improve device efficiency are well documented within the literature, it is often difficult to understand how methods change the electrical properties of the materials. To elucidate this, we fabricated four organic solar cells, two made using the P3HT:PCBM material system and two made from the PTB7:PCBM system (details of the experimental method can be found in the appendix). One of the P3HT:PCBM cells was annealed, while the other was left as cast. For the PTB7:PCBM cells one was cast with +%3 of the additive diiodooctane (DIO) in the solvent of the active layer, while the other had no DIO present. Figure 6, plots the dark JV, light JV, and Suns-V oc curves for the fabricated cells.
It can be seen that the P3HT:PCBM cell which was annealed performed better than the as cast cell, while the PTB7:PCBM cell with DIO added performed better that the one without DIO. Table 4 plots the device parameters extracted from the P3HT:PCBM data set using a trained neural network. The value colored green on each row represents the parameter which would be expected to result in a more efficient device, the one Adv. Funct. Mater. 2020, 30,1907259 Table 2. Material parameters the neural network extracted from the dark and light JV curves for thick (350 nm) and thin (50 nm) devices plotted in Figure 2 for the PBTZT-stat-BDTT-8:PC 61 BM material system.  colored red represents the value which would be expected to result in a less efficient device. It can be seen that annealing has increased the free carrier mobility (µ avg free ), unchanged the energetic disorder (E avg U ), increased the mobility at the maximum power point (µ avg max P ) and increased the recombination constant at the maximum power point ( P avg max τ ). Annealing also appears to have reduced the shunt resistance, which can also be seen from the figure directly (more current at low voltages), and increased the contact resistance. Table 5, plots the device parameters extracted by a neural network from the PTB7:PCBM cells depicted in Figure 6. It can be seen that, the cell with the additive has a higher free carrier mobility (µ avg free ) and a higher mobility at the maximum power point (µ avg max P ). The recombination time constant ( P avg max τ ) was about the same for both devices and contact resistance also slightly worse after the addition of DIO.
Finally, it should be noted that neural networks are not the only machine learning method that can be applied to this data set to extract material parameters. In the Supporting Information we apply a k-nearest neighbor regression to the data set to produce a duplicate of Table 1, with the results being only marginally worse than those produced by the neural network. We chose neural networks for this work because they are arguably currently producing some of the most exciting advances in machine learning.

Conclusion
We have demonstrated that a deep neural network can be used to quickly and robustly extract material parameters from experimental data, reducing the time for device/material parameter extraction from over a week using traditional modeling techniques to seconds. We have verified the results against a more traditional strategy of fitting a device model to an experimental data set. The advantage of using a deep neural network over more traditional approaches, such as fitting a model or analytical expression is threefold: a) once trained the network requires little interaction from the lab scientist to use; b) its use is instantaneous, meaning that values of mobility, trap density, and recombination rates can all be calculated as soon as the experiment is over; and c) unlike the application of methods such as CELIV there are no assumptions made in the application of the neural network. We anticipate this work will be useful for those searching for more efficient third generation solar cell materials. Although, we applied this method to solar cells within the scope of this paper, the method is equally applicable to other classes of devices such as sensors, light emitting diodes or transistors, as long as one can reliably simulate a device using a device model, a neural network can be trained to extract meaning from the experimental data.