Patterned voltage pulse trains are commonly used in neuroscience research to precisely control stimulus isolators (Flaherty and Graybiel, 1994; Bisley et al., 2001; Cohen and Newsome, 2004; Histed et al., 2009), light sources for optogenetic manipulations (Boyden et al., 2005; Cardin et al., 2009), sensory stimuli (Soto-Faraco et al., 2002), and to synchronize events between instruments (Nikolic et al., 2009). Pulse trains can also be also triggered by a particular experimental contingency, providing closed-loop feedback at low latency (Girardeau et al., 2009; Venkatraman et al., 2009; Berényi et al., 2012; Newman et al., 2013). Laboratory instruments specialized for these purposes are commercially available, for example Master 8 (AMPI), PSG-2 (ISSI), Pulsemaster A300 (WPI), BPG-1 (Bak Electronics), StimPulse PGM (FHC Inc.) and Multistim 3800 (A-M Systems). Commercial solutions have been widely adopted, however their cost is a constraint in research and educational settings with limited funding. As proprietary instruments, researchers are also unable to add hardware or software features that would suit their unique needs: for instance, to implement a custom set of triggering rules in firmware, or to develop an interface to the device in a desired programming language. This flexibility can be especially beneficial for experimental design in systems neuroscience, where integration of custom instrumentation is frequently employed to measure and control behavior (Brunton et al., 2013), acquire neural data (Yamamoto and Wilson, 2008; Karlsson and Frank, 2009) and to stimulate the brain both electrically (O’Doherty et al., 2009) and optically (O’Connor et al., 2013).

To address these needs, we developed Pulse Pal (Figure 1), an open source pulse train generator costing ~$210 (US) in easily obtained parts, with essential functionality comparable to commercial stimulators.

To validate Pulse Pal as a practical solution for stimulus control, we tested the precision and reliability of the shortest pulses the device can process on both trigger and output channels, and several other properties relevant for neurophysiology research. All tests were performed on a single Pulse Pal device, connected to a controlling computer (Macbook Pro, Apple).

While output channels are updated once per execution of the microcontroller’s 50 µs main loop, the shortest configurable pulse is restricted to 100 µs (to ensure that the smallest output channel pulse can also reliably trigger the device). To measure the precision of a 100 µs pulse, we programmed Pulse Pal to deliver a train of three 100 µs pulses, separated by 100 µs intervals on all 4 output channels each time a software trigger was detected. Pulse Pal’s first output channel was connected to a digital oscilloscope (DS1102D, Rigol). Pulse Pal was then software-triggered 100,000 times over 24 h by a custom test script written in MATLAB r2013a (Mathworks) on the controlling computer. After each trigger event, the resulting waveform was returned from the oscilloscope to the computer. Pulse Pal generated a unique waveform of three pulses after each trigger, demonstrating high software trigger reliability. The first 100 pulse trains are shown superimposed in Figure 4A, aligned to first pulse onset to demonstrate pulse jitter. The pulse widths of all 300,000 pulses are shown in Figure 4B. Cycle widths ranged from 96.9 µs to 102.9 µs, and 99.97% of pulses were within 3 µs of 100 µs.

Next, we measured clock drift, to ensure that the timing variability we observed in 100 µs-long pulses did not propagate. We programmed Pulse Pal to deliver a single 10 s pulse (spanning 200,000 microcontroller loop cycles) when triggered. We captured the resulting waveform on 10,000 trials with an analog capture device (NI USB-6210, National Instruments), sampled at 100 kHz. 100 example pulse waveforms are shown in Figure 4C, and all 10,000 pulse widths are shown in Figure 4D. All pulses measured either 9.99998 s or 9.99997 s, corresponding to a consistent clock drift of 3 µs/s with respect to the NI USB-6210 clock.

To measure pulse train latency and trigger channel reliability, we connected two Pulse Pals in series. The first was soft-triggered by the computer on each of 100,000 trials. It generated a single 5 V, 100 µs square pulse simultaneously on two output channels—one delivered to the first trigger channel of the second Pulse Pal, and the other to an oscilloscope (see Figure 4H for a separate experiment demonstrating the simultaneity of these pulses). The second Pulse Pal generated a single 100 µs pulse on each output channel when its first trigger channel was triggered, which was captured from one output channel by a separate oscilloscope channel. In Figure 4E, 100 trials are shown. Trigger pulses from the first (triggering) Pulse Pal are shown in red, and pulses from the second Pulse Pal in black. All 100,000 pulses captured were unique waveforms, indicating high hardware trigger reliability. Output channel latency for all 100,000 trials ranged relatively uniformly between 91.0 and 146.7 µs (Figure 4F).

Next we sought to determine whether channel noise was low enough that a single bit DAC increment produced a non-overlapping change in voltage. Since the MAX500ACPE+ DAC that drives Pulse Pal’s output channels has 8-bits of precision mapped across a 20 V range (−10 V to +10 V), its least significant bit (LSB) increments the channel voltage by 78.1 mV. Therefore we set up PulsePal to trigger two 78.1 mV pulses 100 times (shown in Figure 4G). Fluctuations in voltage on individual trials ranged as much as 8 mV about the mean in the 100 ms interval prior to pulse train onset, and as much as 65 mV about the mean during the first 100 ms pulse. The increased noise during playback was mostly attributable to digital feed-through from the SPI channels controlling the DAC (data not shown), but remained significantly less than the DAC’s minimum voltage increment. While several board layout and circuit modifications could be implemented by researchers wishing to further reduce digital feed-through (for instance, by optical isolation of the SPI bus), Pulse Pal can exploit the full bit-width of its DAC in its present form, rendering it useful for many control applications in neuroscience instrumentation. The resting voltage of the output channel was programmatically set to 0 V, but was measured in this experiment to be 10.55 mV (within the 15 mV “zero code error” specified for the MAX500ACPE+ DAC in its datasheet), indicating that an offset from the 0 V set-point was present but slight.

In experiments with precisely timed events, it is useful to produce signals that occur simultaneously. Therefore we measured the simultaneity of output channel updates by comparing pulses triggered on the first and last output channels. We set the first and fourth output channels to a resting voltage of −10 V, delivered 100 +10 V pulses, and captured the rising waveform of each pulse with the oscilloscope (Figure 4H). On all trials, the output voltage on both channels settled within 100 mV of +10 V after 3.5 µs. This measurement also confirmed that DAC and output amplifier slew rates were fast enough to produce 100 us pulses useful for most applications in neuroscience research.

In many experiments, pulse train parameters and timing data must be updated rapidly in response to recently acquired information. Since the same microcontroller controls pulse timing and USB communication, Pulse Pal cannot be updated while a pulse train is being delivered. Therefore we sought to ensure that updates could be achieved rapidly between experimental trials. We measured the USB data transfer speed by sending a 1000-pulse train (5006 bytes) 100 times, from the Pulse Pal MATLAB client to Pulse Pal. For a measure of the hardware transfer speed without client-side software overhead, Pulse Pal’s firmware was modified to indicate the start of data transfer by setting output channel 1–5 V, and the end of data transfer by returning channel 1–0 V. The resulting pulse was captured by the oscilloscope on each trial. The client-side transfer time was separately measured for the blocking MATLAB serial fwrite command, by flanking it with tic and toc commands. Transfers completed in 26–35 ms (averaging 171 KB/s), while client-side overhead cost an additional 12 ms on average (Figure 4I). Consistent with this transfer speed measurement, a separate transfer updating all of Pulse Pal’s channel parameters for all channels (163 bytes) completed on the hardware side in less than 1 ms (data not shown).

Finally, we sought to verify Pulse Pal’s suitability for precise optical control in optogenetic experiments, by using it to control a 447 nm diode laser’s timing and intensity. We connected the laser (BML447-50FLD, Lasermate Group) through an optical fiber (M31L02, ThorLabs) to a silicon transimpedence amplified photodetector (PDA10A, ThorLabs), delivered 1 ms pulses from a Pulse Pal output channel to the laser power supply’s analog input, and captured the resulting waveforms with an oscilloscope (DS1102D, Rigol). Pulses ranged in amplitude from 78 mV to 5.0 V in 78 mV increments. In Figure 4J, the single traces captured for each voltage are shown superimposed, skipping every second voltage for clarity. Pulse Pal elicited precisely gated pulses of light from the laser, with programmatic (albeit slightly nonlinear) control of light intensity.

For our research, we required a pulse generator with high precision at timescales relevant for alignment of stimulation events to action potentials (pulse time jitter at least a factor of 10 less than an action potential width; Figures 4A,B), low clock drift (Figures 4C,D), first pulse latency (Figures 4E,F) comparable to short mammalian action potentials (Kandel et al., 2000) and high reliability (100% of 300,000 soft triggers, 100% of 100,000 100 µs trigger pulses). In developing Pulse Pal, we recognized that simplified control of voltage pulse timing is a general need, and expanded the project’s scope to meet five additional design objectives: low material cost ($210 USD), stand-alone functionality (Figure 1), bench-side assembly with common tools (see illustrated guide on wiki), support for common computing platforms and programming languages (WinXP, Win7, OSX, Ubuntu 14.04; MATLAB, C++, Python) and comprehensive online documentation.

In fulfilling these objectives, Pulse Pal provides a general resource for precise temporal control of stimulation and environmental cues in the laboratory. It encapsulates the problem of temporal pattern generation for many applications in physiology and psychophysics, where in lieu of commercial instrumentation, this control problem had often been addressed ad-hoc by writing custom software for microcontrollers (da Silva Pinto et al., 2011; Weick et al., 2011; Bugaj et al., 2013; Haikala et al., 2013; Ohayon et al., 2013; Smear et al., 2013; Inagaki et al., 2014; Klapoetke et al., 2014).

Pulse Pal’s parametric approach to stimulation features trigger logic rules and stimulus pattern motifs commonly implemented in neuroscience research, however custom applications may require the device to perform less common functions. Unlike its commercial counterparts, Pulse Pal’s firmware is provided in the public domain with an open source license. To facilitate access, the firmware was written in the Arduino language⁴ —a reduced set of C++ syntax with extensive online documentation for developers who lack a programming background. The adaptation of Arduino for Pulse Pal’s microcontroller platform⁵ exposes further functionality specific to the ARM Cortex M3 microcontroller. We anticipate that this lower barrier to entry will be exploited by researchers using Pulse Pal’s hardware, firmware and software as a starting point for tailored applications beyond its present niche.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.