Signal-space separation (SSS) and Maxwell filtering#

This tutorial covers reducing environmental noise and compensating for head movement with SSS and Maxwell filtering.

As usual, we’ll start by importing the modules we need, loading some example data, and cropping it to save on memory:

# License: BSD-3-Clause
# Copyright the MNE-Python contributors.
import os

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns

import mne
from mne.preprocessing import find_bad_channels_maxwell

sample_data_folder = mne.datasets.sample.data_path()
sample_data_raw_file = os.path.join(
    sample_data_folder, "MEG", "sample", "sample_audvis_raw.fif"
raw =, verbose=False)
Measurement date December 03, 2002 19:01:10 GMT
Experimenter MEG
Participant Unknown
Digitized points 146 points
Good channels 203 Gradiometers, 102 Magnetometers, 9 Stimulus, 59 EEG, 1 EOG
Bad channels MEG 2443, EEG 053
EOG channels EOG 061
ECG channels Not available
Sampling frequency 600.61 Hz
Highpass 0.10 Hz
Lowpass 172.18 Hz
Projections PCA-v1 : off
PCA-v2 : off
PCA-v3 : off
Filenames sample_audvis_raw.fif
Duration 00:01:01 (HH:MM:SS)

Background on SSS and Maxwell filtering#

Signal-space separation (SSS) [1][2] is a technique based on the physics of electromagnetic fields. SSS separates the measured signal into components attributable to sources inside the measurement volume of the sensor array (the internal components), and components attributable to sources outside the measurement volume (the external components). The internal and external components are linearly independent, so it is possible to simply discard the external components to reduce environmental noise. Maxwell filtering is a related procedure that omits the higher-order components of the internal subspace, which are dominated by sensor noise. Typically, Maxwell filtering and SSS are performed together (in MNE-Python they are implemented together in a single function).

Like SSP, SSS is a form of projection. Whereas SSP empirically determines a noise subspace based on data (empty-room recordings, EOG or ECG activity, etc) and projects the measurements onto a subspace orthogonal to the noise, SSS mathematically constructs the external and internal subspaces from spherical harmonics and reconstructs the sensor signals using only the internal subspace (i.e., does an oblique projection).


Maxwell filtering was originally developed for Elekta Neuromag® systems, and should be considered experimental for non-Neuromag data. See the Notes section of the maxwell_filter() docstring for details.

The MNE-Python implementation of SSS / Maxwell filtering currently provides the following features:

  • Basic bad channel detection (find_bad_channels_maxwell())

  • Bad channel reconstruction

  • Cross-talk cancellation

  • Fine calibration correction

  • tSSS

  • Coordinate frame translation

  • Regularization of internal components using information theory

  • Raw movement compensation (using head positions estimated by MaxFilter)

  • cHPI subtraction (see mne.chpi.filter_chpi())

  • Handling of 3D (in addition to 1D) fine calibration files

  • Epoch-based movement compensation as described in [1] through mne.epochs.average_movements()

  • Experimental processing of data from (un-compensated) non-Elekta systems

Using SSS and Maxwell filtering in MNE-Python#

For optimal use of SSS with data from Elekta Neuromag® systems, you should provide the path to the fine calibration file (which encodes site-specific information about sensor orientation and calibration) as well as a crosstalk compensation file (which reduces interference between Elekta’s co-located magnetometer and paired gradiometer sensor units).

fine_cal_file = os.path.join(sample_data_folder, "SSS", "sss_cal_mgh.dat")
crosstalk_file = os.path.join(sample_data_folder, "SSS", "ct_sparse_mgh.fif")

Before we perform SSS we’ll look for bad channels — MEG 2443 is quite noisy.


It is critical to mark bad channels in['bads'] before calling maxwell_filter() in order to prevent bad channel noise from spreading.

Let’s see if we can automatically detect it.["bads"] = []
raw_check = raw.copy()
auto_noisy_chs, auto_flat_chs, auto_scores = find_bad_channels_maxwell(
print(auto_noisy_chs)  # we should find them!
print(auto_flat_chs)  # none for this dataset
Applying low-pass filter with 40.0 Hz cutoff frequency ...
Reading 0 ... 36037  =      0.000 ...    60.000 secs...
Filtering raw data in 1 contiguous segment
Setting up low-pass filter at 40 Hz

FIR filter parameters
Designing a one-pass, zero-phase, non-causal lowpass filter:
- Windowed time-domain design (firwin) method
- Hamming window with 0.0194 passband ripple and 53 dB stopband attenuation
- Upper passband edge: 40.00 Hz
- Upper transition bandwidth: 10.00 Hz (-6 dB cutoff frequency: 45.00 Hz)
- Filter length: 199 samples (0.331 s)

[Parallel(n_jobs=1)]: Done  17 tasks      | elapsed:    0.0s
[Parallel(n_jobs=1)]: Done  71 tasks      | elapsed:    0.1s
[Parallel(n_jobs=1)]: Done 161 tasks      | elapsed:    0.3s
[Parallel(n_jobs=1)]: Done 287 tasks      | elapsed:    0.5s
Scanning for bad channels in 12 intervals (5.0 s) ...
    No bad MEG channels
    Processing 204 gradiometers and 102 magnetometers
    Using fine calibration sss_cal_mgh.dat
        Adjusting non-orthogonal EX and EY
        Adjusted coil positions by (μ ± σ): 0.5° ± 0.4° (max: 2.1°)
    Automatic origin fit: head of radius 91.0 mm
    Using origin -4.1, 16.0, 51.8 mm in the head frame
        Interval   1:    0.000 -    4.998
        Interval   2:    5.000 -    9.998
        Interval   3:   10.000 -   14.998
        Interval   4:   15.000 -   19.998
        Interval   5:   20.000 -   24.998
        Interval   6:   24.999 -   29.998
        Interval   7:   29.999 -   34.997
        Interval   8:   34.999 -   39.997
        Interval   9:   39.999 -   44.997
        Interval  10:   44.999 -   49.997
        Interval  11:   49.999 -   54.997
        Interval  12:   54.999 -   60.000
    Static bad channels:  ['MEG 2443']
    Static flat channels: []
['MEG 2443']


find_bad_channels_maxwell needs to operate on a signal without line noise or cHPI signals. By default, it simply applies a low-pass filter with a cutoff frequency of 40 Hz to the data, which should remove these artifacts. You may also specify a different cutoff by passing the h_freq keyword argument. If you set h_freq=None, no filtering will be applied. This can be useful if your data has already been preconditioned, for example using mne.chpi.filter_chpi(),, or

Now we can update the list of bad channels in the dataset.

We called find_bad_channels_maxwell with the optional keyword argument return_scores=True, causing the function to return a dictionary of all data related to the scoring used to classify channels as noisy or flat. This information can be used to produce diagnostic figures.

In the following, we will generate such visualizations for the automated detection of noisy gradiometer channels.

# Only select the data for gradiometer channels.
ch_type = "grad"
ch_subset = auto_scores["ch_types"] == ch_type
ch_names = auto_scores["ch_names"][ch_subset]
scores = auto_scores["scores_noisy"][ch_subset]
limits = auto_scores["limits_noisy"][ch_subset]
bins = auto_scores["bins"]  # The the windows that were evaluated.
# We will label each segment by its start and stop time, with up to 3
# digits before and 3 digits after the decimal place (1 ms precision).
bin_labels = [f"{start:3.3f}{stop:3.3f}" for start, stop in bins]

# We store the data in a Pandas DataFrame. The seaborn heatmap function
# we will call below will then be able to automatically assign the correct
# labels to all axes.
data_to_plot = pd.DataFrame(
    columns=pd.Index(bin_labels, name="Time (s)"),
    index=pd.Index(ch_names, name="Channel"),

# First, plot the "raw" scores.
fig, ax = plt.subplots(1, 2, figsize=(12, 8), layout="constrained")
    f"Automated noisy channel detection: {ch_type}", fontsize=16, fontweight="bold"
sns.heatmap(data=data_to_plot, cmap="Reds", cbar_kws=dict(label="Score"), ax=ax[0])
    ax[0].axvline(x, ls="dashed", lw=0.25, dashes=(25, 15), color="gray")
    for x in range(1, len(bins))
ax[0].set_title("All Scores", fontweight="bold")

# Now, adjust the color range to highlight segments that exceeded the limit.
    vmin=np.nanmin(limits),  # bads in input data have NaN limits
    ax[1].axvline(x, ls="dashed", lw=0.25, dashes=(25, 15), color="gray")
    for x in range(1, len(bins))
ax[1].set_title("Scores > Limit", fontweight="bold")
Automated noisy channel detection: grad, All Scores, Scores > Limit


You can use the very same code as above to produce figures for flat channel detection. Simply replace the word “noisy” with “flat”, and replace vmin=np.nanmin(limits) with vmax=np.nanmax(limits).

You can see the un-altered scores for each channel and time segment in the left subplots, and thresholded scores – those which exceeded a certain limit of noisiness – in the right subplots. While the right subplot is entirely white for the magnetometers, we can see a horizontal line extending all the way from left to right for the gradiometers. This line corresponds to channel MEG 2443, which was reported as auto-detected noisy channel in the step above. But we can also see another channel exceeding the limits, apparently in a more transient fashion. It was therefore not detected as bad, because the number of segments in which it exceeded the limits was less than 5, which MNE-Python uses by default.


You can request a different number of segments that must be found to be problematic before find_bad_channels_maxwell reports them as bad. To do this, pass the keyword argument min_count to the function.

Obviously, this algorithm is not perfect. Specifically, on closer inspection of the raw data after looking at the diagnostic plots above, it becomes clear that the channel exceeding the “noise” limits in some segments without qualifying as “bad”, in fact contains some flux jumps. There were just not enough flux jumps in the recording for our automated procedure to report the channel as bad. So it can still be useful to manually inspect and mark bad channels. The channel in question is MEG 2313. Let’s mark it as bad:["bads"] += ["MEG 2313"]  # from manual inspection

After that, performing SSS and Maxwell filtering is done with a single call to maxwell_filter(), with the crosstalk and fine calibration filenames provided (if available):

raw_sss = mne.preprocessing.maxwell_filter(
    raw, cross_talk=crosstalk_file, calibration=fine_cal_file, verbose=True
Maxwell filtering raw data
    Bad MEG channels being reconstructed: ['MEG 2443', 'MEG 2313']
    Processing 204 gradiometers and 102 magnetometers
    Using fine calibration sss_cal_mgh.dat
        Adjusting non-orthogonal EX and EY
        Adjusted coil positions by (μ ± σ): 0.5° ± 0.4° (max: 2.1°)
    Automatic origin fit: head of radius 91.0 mm
    Using origin -4.1, 16.0, 51.8 mm in the head frame
        Using 87/95 harmonic components for    0.000  (72/80 in, 15/15 out)
    Loading raw data from disk
    Processing 6 data chunks

To see the effect, we can plot the data before and after SSS / Maxwell filtering.

raw.pick(["meg"]).plot(duration=2, butterfly=True)
raw_sss.pick(["meg"]).plot(duration=2, butterfly=True)
  • Raw plots
  • Raw plots

Notice that channels marked as “bad” have been effectively repaired by SSS, eliminating the need to perform interpolation. The heartbeat artifact has also been substantially reduced.

The maxwell_filter() function has parameters int_order and ext_order for setting the order of the spherical harmonic expansion of the interior and exterior components; the default values are appropriate for most use cases. Additional parameters include coord_frame and origin for controlling the coordinate frame (“head” or “meg”) and the origin of the sphere; the defaults are appropriate for most studies that include digitization of the scalp surface / electrodes. See the documentation of maxwell_filter() for details.

Spatiotemporal SSS (tSSS)#

An assumption of SSS is that the measurement volume (the spherical shell where the sensors are physically located) is free of electromagnetic sources. The thickness of this source-free measurement shell should be 4-8 cm for SSS to perform optimally. In practice, there may be sources falling within that measurement volume; these can often be mitigated by using Spatiotemporal Signal Space Separation (tSSS) [2]. tSSS works by looking for temporal correlation between components of the internal and external subspaces, and projecting out any components that are common to the internal and external subspaces. The projection is done in an analogous way to SSP, except that the noise vector is computed across time points instead of across sensors.

To use tSSS in MNE-Python, pass a time (in seconds) to the parameter st_duration of maxwell_filter(). This will determine the “chunk duration” over which to compute the temporal projection. The chunk duration effectively acts as a high-pass filter with a cutoff frequency of \(\frac{1}{\mathtt{st\_duration}}~\mathrm{Hz}\); this effective high-pass has an important consequence:

  • In general, larger values of st_duration are better (provided that your computer has sufficient memory) because larger values of st_duration will have a smaller effect on the signal.

If the chunk duration does not evenly divide your data length, the final (shorter) chunk will be added to the prior chunk before filtering, leading to slightly different effective filtering for the combined chunk (the effective cutoff frequency differing at most by a factor of 2). If you need to ensure identical processing of all analyzed chunks, either:

  • choose a chunk duration that evenly divides your data length (only recommended if analyzing a single subject or run), or

  • include at least 2 * st_duration of post-experiment recording time at the end of the Raw object, so that the data you intend to further analyze is guaranteed not to be in the final or penultimate chunks.

Additional parameters affecting tSSS include st_correlation (to set the correlation value above which correlated internal and external components will be projected out) and st_only (to apply only the temporal projection without also performing SSS and Maxwell filtering). See the docstring of maxwell_filter() for details.

Movement compensation#

If you have information about subject head position relative to the sensors (i.e., continuous head position indicator coils, or cHPI), SSS can take that into account when projecting sensor data onto the internal subspace. In case you are not sure whether you have that information or want to doublecheck, you can use mne.chpi.get_active_chpi() (currently only implemented for neuromag systems).

chpi_fif_file = os.path.join(
    mne.datasets.testing.data_path(), "SSS", "test_move_anon_raw.fif"
raw =, allow_maxshield="yes")

# time-resolved information on active HPI coils
# if all hpi were inactive n_active is a zero-array
n_active = mne.chpi.get_active_chpi(raw)
print(f"Average number of coils active during recording: {n_active.mean()}")
Opening raw data file /home/circleci/mne_data/MNE-testing-data/SSS/test_move_anon_raw.fif...
    Read a total of 12 projection items:
        mag.fif : PCA-v1 (1 x 306)  idle
        mag.fif : PCA-v2 (1 x 306)  idle
        mag.fif : PCA-v3 (1 x 306)  idle
        mag.fif : PCA-v4 (1 x 306)  idle
        mag.fif : PCA-v5 (1 x 306)  idle
        mag.fif : PCA-v6 (1 x 306)  idle
        mag.fif : PCA-v7 (1 x 306)  idle
        grad.fif : PCA-v1 (1 x 306)  idle
        grad.fif : PCA-v2 (1 x 306)  idle
        grad.fif : PCA-v3 (1 x 306)  idle
        grad.fif : PCA-v4 (1 x 306)  idle
        Average EEG reference (1 x 60)  idle
    Range : 10800 ... 31199 =      9.000 ...    25.999 secs
Using 5 HPI coils: 83 143 203 263 323 Hz
Average number of coils active during recording: 5.0

Head position data can be computed using mne.chpi.compute_chpi_locs() and mne.chpi.compute_head_pos(), or loaded with the mne.chpi.read_head_pos() function. The example data doesn’t include cHPI, so here we’ll load a .pos file used for testing, just to demonstrate:

head_pos_file = os.path.join(
    mne.datasets.testing.data_path(), "SSS", "test_move_anon_raw.pos"
head_pos = mne.chpi.read_head_pos(head_pos_file)
mne.viz.plot_head_positions(head_pos, mode="traces")
Position (mm), Rotation (quat)

The cHPI data file could also be passed as the head_pos parameter of maxwell_filter(). Not only would this account for movement within a given recording session, but also would effectively normalize head position across different measurement sessions and subjects. See here for an extended example of applying movement compensation during Maxwell filtering / SSS. Another option is to apply movement compensation when averaging epochs into an Evoked instance, using the mne.epochs.average_movements() function.

Each of these approaches requires time-varying estimates of head position, which is obtained from MaxFilter using the -headpos and -hp arguments (see the MaxFilter manual for details).

Caveats to using SSS / Maxwell filtering#

  1. There are patents related to the Maxwell filtering algorithm, which may legally preclude using it in commercial applications. More details are provided in the documentation of maxwell_filter().

  2. SSS works best when both magnetometers and gradiometers are present, and is most effective when gradiometers are planar (due to the need for very accurate sensor geometry and fine calibration information). Thus its performance is dependent on the MEG system used to collect the data.


Total running time of the script: (0 minutes 18.433 seconds)

Estimated memory usage: 160 MB

Gallery generated by Sphinx-Gallery