Data

Typical manual-tuning-inspired measurements are one- and two-dimensional traces on a grid. While this is fairly limiting when dealing with multi-dimensional spaces, it is currently the only human-analyzable way of taking data. As nanotune’s tuning sequence is based on existing manual procedures, the measurements it supports are 1D and 2D traces.

All measurements are done using QCoDeS and are thus saved to a database via a QCoDeS dataset. However, nanotune has its own dataset abstraction with a few extra functionalities convenient for tuning.

Dataset

The dataset class emulates a QCoDeS dataset in order to standardize a few naming conventions and adding post-processing steps. Using QCoDeS’s to_xarray_dataset, raw data is loaded into an xarray called raw_data. Using the normalization constants saved to metadata of the QCoDeS dataset, the raw data is normalized and retained in the data attribute, also an xarray dataset. The variables of this xarray are renamed to match the standardized readout methods ‘transport’, ‘sensing’ and ‘rf’. The reason being that tuning and classification procedures need to know which trace belongs to which readout methods. The information which readout instruments corresponds to which readout method is saved to metadata during measurements. Similarly, xarray datasets with a filtered signal and the power spectrum are computed as well.

In general, all data measured with QCoDeS can be loaded into a nanotune dataset. However, data not measured using nanotune will miss some metadata and thus its normalization will be incorrect. The required metadata is saved into QCoDeS’ snapshot under the nt.meta_tag key, which is defined in nanotune.configuration.config.json.

The following code is an example of how to load data into a nanotune dataset.

import os
import nanotune as nt

nt_root = os.path.dirname(os.path.dirname(os.path.abspath(nt.__file__)))
db_pinchoff = os.path.join('data', 'tuning', 'device_characterization.db')

ds = nt.Dataset(1203, db_pinchoff, db_folder=nt_root)

For more details see also the dataset example notebook,

Databases

As shown above and as opposed to QCoDes, nanotune does not rely on data run IDs only. It also needs the database name and optionally the folder where the database is located to find data. The reason is that the volume of data recorded during automated tuning grows significantly faster than during manual tuning. Not only becomes copying databases larger than 10Gb cumbersome, but sometimes one wishes to save data to several databases. One example are parallel measurements, for which one database per sample could be used to avoid locked .db files.

To facilitate switching databases as well as other tuning/classification related functionalities, nanotune has a few helper functions. The most used one simply sets a different database as the default accessed by QCoDeS and nanotune:

import nanotune as nt
nt.set_database("examples.db", db_folder="./")

Further, there are also functions searching for data with a specific machine learning label, e.g. good pinchoffs. As these labels occupy a separate column in the database, these functions utilize efficient SQLite queries. The columns are initialized whenever a database is created using ‘nt.new_database’.

Data flow

The diagram shown in Fig. 13 illustrates the data and instruction flow of quantum measurements. The dotted ellipses indicates which stages are covered by QCoDeS or nanotune. QCoDeS provides an interface to room-temperature instruments (i.e. drivers) and tools to take and save measurements, while nanotune extends this functionality by automating common procedures encountered during quantum dot initialization.

As stated above, all data is taken via QCoDeS, specifically by making use of its measurement context manager. The data is taken during the device characterization and dot-tuning sequence, which is implemented by the Device tuner classes Tuner, Characterizer and DotTuner. Specifically, the tuner classes call Tuningstages subclasses, i.e. GateCharacterization1D and ChargeDiagram, which are responsible for measurements and correct metadata saving.

Once measured, the data is loaded into nanotune’s dataset, where the raw data is post-processed, e.g. normalized. If required, Fourier frequencies or filtered data is computed as well. The DataFit classes then extract features such as the slope and amplitude of a pinchoff curve. Based on these features as well as min and max values of the measured signal, the device’s transport regime, i.e. open, intermediate or closed, can be determined. Either the extracted feature vector or entire measurement is passed to the classifier for quality or charge state prediction. Based on the outcome, a decision about subsequent tuning is made.