Improving the time-resolution of event-based electron cameras using neural networks trained by femtosecond electron pulse detection

Abstract text (incl. figure legends and references)

Introduction

In the recent two decades, ultrafast transmission electron microscopy (UTEM) has successfully enabled investigation of ultrafast dynamics on the nanoscale by combining femtosecond electron probe pulses with a synchronized optical sample excitation [1]. With increasing availability of sufficiently fast event-based electron detectors, dynamics down to the nanosecond timescale can also be imaged by continuous-beam TEM. For example, with the TimePix3 chip architecture [2] electron events are registered with a bin width of about 1.6 ns. However, the achievable spatial and temporal resolution in experiments is often limited due to the stochastic electron trajectories in the sensor chip [3]. Neural network approaches have already enabled increased spatial resolution [4].

Objectives

Here, we report on the development of a neural network architecture trained by experimental electron pulse data to improve the temporal resolution of the TimePix3 detector.

Materials & methods

As a high-precision electron arrival time standard, we utilized 200-fs electron pulses (400-kHz repetition rate, 200-keV electron energy, 1.3 electrons/pulse) using the Oldenburg UTEM (Fig. 1a). With a TimePix3 detector (Cheetah T3, Amsterdam Scientific Instruments), we experimentally collected a dataset of electron detection events for 4x10⁶ electron pulses. For accurate synchronization, the TimePix3 detector additionally assigns timestamps to each photoemission laser pulse. For every event, relative detector position, time of arrival (ToA) and time over threshold (ToT) are stored and clustered according to their arrival time and position (on average 6 events per cluster). To predict the arrival time of the primary electron, a fully connected deep neural network model is trained using event data from 1.6x10⁵ time-scrambled event cluster.

Results

The unprocessed ToA data, using individual unclustered event data, shows a temporal spread of about 10 ns FWHM (Fig. 1b (gray)). Already by identifying event clusters and considering the average ToA in each cluster, the temporal spread in the histogram can be reduced to 7 ns (red). With the neural network, the predicted ToA distribution (for a part of the data not involved in the training) is narrowed to 4.7 ns (FWHM, 2 ns rms, blue). Remarkably, the tailed ToA distribution in the previous cases is transformed to an almost Gaussian distribution by the neural network. An example of imaging data for the unprocessed and neural-network-filtered case is shown in Fig. 1c and d, respectively.

Conclusion

In conclusion, we demonstrate the application of neural networks trained by femtosecond electron pulse data to predict the accurate ToA, substantially increasing the temporal resolution of the TimePix3 detector. We expect that future versions of the TimePix chip with shorter time bins will likely benefit even more from sophisticated event data analysis.

[1] A. Feist et al., Ultramicroscopy 176, 63 (2017).

[2] T. Poikela et al., JINST 9, C05013 (2014).

[3] F. Pitters et al., JINST 14, P05022 (2019).

[4] J. P. van Schayck et al., Ultramicroscopy 218, 113091 (2020).

Figure 1: (a) Schematic setup of Oldenburg UTEM with TimePix3 detector. (b) Histogram of the time of arrival for uncorrected (gray), cluster mean corrected ToA (red), and the neural network predicted ToA (blue). (c,d) Image for uncorrected (c) and neural network predicted ToA (d). Color scale: ToA of electron event.